The appeal of synthetic CT, or sCT, lies in a fundamental limitation of cone-beam CT acquired with medical linear accelerators. CBCT can show the anatomy of a patient during a treatment fraction, making it useful for positioning and adaptive radiotherapy, but scatter, beam hardening, noise, metal artifacts and weak soft-tissue contrast can distort the Hounsfield unit values used to infer tissue density. An sCT model attempts to correct those defects while retaining the fraction-specific anatomy visible in CBCT. The resulting image is intended not merely to look like conventional CT, but to provide quantitative information that can support electron-density estimation, dose recalculation, treatment-plan evaluation and adaptation decisions.

The review organizes the field into four overlapping methodological families. Supervised deterministic regression, commonly based on convolutional neural networks and U-Net architectures, learns a direct mapping from CBCT to CT-equivalent images. Adversarial systems such as Pix2Pix and CycleGAN use discriminator-based training to generate images resembling the CT domain. Registration-aware approaches explicitly model spatial mismatch between CBCT and a reference CT, while probabilistic diffusion models learn to reconstruct an image through an iterative, CBCT-conditioned denoising process. Hybrid systems combine these elements, and their defining feature determines where the review places them. No method is consistently superior across every anatomical region, imaging device and treatment modality.

Supervised regression generally provides stable optimization and relatively rapid inference when well-aligned CBCT–CT pairs are available. CNNs and U-Nets can reduce intensity bias and improve Hounsfield-unit consistency, while additional contextual or feature-based losses may help preserve tissue boundaries. Their weakness is that voxel-wise objectives can average several plausible solutions and produce overly smooth images, and skip connections can carry input artifacts into the output. Transformers and vision transformers can capture long-range spatial relationships, including in a three-dimensional head-and-neck proton therapy study, but require greater data, memory and computing resources. Registration errors remain a fundamental problem because anatomical change, positioning differences and organ filling can be learned as false intensity relationships.

Adversarial methods may improve texture and local detail beyond direct regression, but they introduce different risks. Pix2Pix combines adversarial learning with a paired reconstruction loss and therefore depends on relatively close spatial correspondence. CycleGAN can learn from unpaired CBCT and CT collections, easing data collection but weakening direct geometric supervision. Such unpaired translation can produce CT-like appearance while altering local structure or inventing features that are not present in the patient. Registration-aware systems such as RegGAN treat residual misalignment as part of the optimization rather than leaving it as an unexplained disturbance. Head-and-neck evaluations have reported improvements in Hounsfield-unit, structural and dosimetric endpoints relative to corresponding generators without explicit registration modeling, although performance still depends on the accuracy of the deformation model and the breadth of the training data.

Diffusion models form one of the four major methodological families examined in the review. They learn a forward noising process and a reverse denoising process, progressively generating an sCT under CBCT guidance without relying on the generator–discriminator min–max optimization used in adversarial systems. A conditional diffusion probabilistic model reportedly outperformed a supervised CycleGAN comparator on image metrics and visual results for paired brain and head-and-neck data under the same registration and evaluation conditions, but that study did not report independent external validation or dosimetric validation. Other work has added frequency, boundary, energy and texture guidance, as well as three-dimensional and hybrid U-Net designs. The main deployment drawback is iterative sampling, which raises inference time and computing cost, although reported durations are difficult to compare because studies measure different units, hardware, image dimensions and numbers of sampling steps.

Image quality alone cannot establish clinical reliability. Common measures such as mean absolute error, peak signal-to-noise ratio, structural similarity, normalized cross-correlation and spatial nonuniformity summarize agreement with a reference CT, but they can conceal clinically important errors in bone, air cavities, soft tissue, tumors or organs at risk. The review therefore recommends region-specific Hounsfield-unit analysis, boundary metrics such as the Dice coefficient and 95th-percentile Hausdorff distance, expert review and patient-level failure analysis. Reference-image quality is also crucial: planning CT may be temporally separated from treatment CBCT, while same-day repeat CT or quality-assurance CT provides a shorter interval. A deformed planning CT is itself a registration-derived reference. These standards are not interchangeable, and a difference between synthetic and reference images may reflect model error, genuine anatomical change, registration failure or a mismatch in acquisition and physiology.

Dosimetric testing asks a different and more demanding question: can the generated image support the treatment task for which it is intended? Photon studies commonly compare dose distributions with the gamma index and dose–volume histograms, along with target and organ-at-risk endpoints. Gamma criteria such as 3%/3 mm, 2%/2 mm and 3%/2 mm are study-specific comparison settings, not universal clinical acceptance thresholds for synthetic CT. A cohort-average gamma pass rate can also conceal an individual patient with unacceptable disagreement. Proton therapy requires additional measures because range is sensitive to tissue composition, air cavities and beam-path composition. Relevant endpoints include relative stopping power, water-equivalent thickness and range error, interpreted alongside plan-dose outcomes. Evidence from photon and proton treatments should therefore remain separate rather than being transferred on the assumption that one modality represents the other.

Translation to adaptive radiotherapy is constrained by more than model inference time. A complete workflow includes data transfer, preprocessing, registration, image generation, contour handling, dose calculation, plan evaluation, quality assurance and recovery when an abnormal result occurs. Selected studies have tested dynamic phantoms or complete retrospective workflows, and one lung workflow reported approximately 11 minutes per treatment fraction on two graphics processors, while another dynamic-thoracic-phantom study reported about 15 seconds for sCT generation but six to eight minutes end to end. Such figures are not directly comparable and remain laboratory or preclinical evidence. The review found little prospective documentation of failures, retries, human overrides, fallback procedures or post-deployment monitoring. It recommends shadow-mode testing, in which outputs are recorded and reviewed without influencing treatment decisions, to determine whether a locked workflow can be monitored and safely controlled.

The review used a structured PubMed search last updated on 30 July 2026, supplemented by reference-list screening and targeted searches, but it was a narrative synthesis rather than a protocol-based systematic review. It did not perform formal risk-of-bias assessment or quantitative meta-analysis, and the included studies varied widely in datasets, scanners, acquisition protocols, reference images, registration methods and evaluation criteria. That heterogeneity makes broad cross-study comparisons uncertain. The authors say the next generation of research should use locked models in independent centers, devices and acquisition protocols; report patient-level structural and Hounsfield-unit errors; preserve modality-specific dosimetric endpoints; and link uncertainty or anomaly detection to predefined quality-assurance triggers, human review and recovery pathways. Until those elements are demonstrated, technically impressive synthetic images should not be confused with evidence that adaptive treatment decisions are safe and reproducible in routine care.

Subject of Research: Clinical validation of deep-learning synthetic CT generated from cone-beam CT for adaptive radiotherapy

Article Title: Cone-beam CT-based synthetic CT generation for adaptive radiotherapy: methodological paradigms, clinical reliability assessment, and translation

Article References: Zhang, Y., Zhang, W., Huang, J., Zhao, J., Luo, H., He, D., & Yang, Z. (2026). Cone-beam CT-based synthetic CT generation for adaptive radiotherapy: methodological paradigms, clinical reliability assessment, and translation. Frontiers in Oncology, 16, Article 1981244. https://doi.org/10.3389/fonc.2026.1981244

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1981244

Keywords: synthetic CT, cone-beam CT, adaptive radiotherapy, deep learning, medical image synthesis, dosimetry, clinical validation, proton therapy, radiotherapy, uncertainty estimation, Cone-beam, CT-based

The challenge is far harder than it might appear. When a clinician photographs a foot, the ulcer may be a barely visible speck against skin that varies enormously in tone, texture, and illumination. The lesion boundary often has low contrast relative to the surrounding tissue, and its shape is wildly irregular, defying the smooth blob-like assumptions that many segmentation networks implicitly make. The researchers behind UlcerNet, based at Vellore Institute of Technology with a collaborator at Deloitte Tohmatsu DeepSquare in Tokyo, note that studies suggest nearly 70 percent of foot ulcers recur even after treatment in at least 10 percent of patients, and that the physical wound also carries a psychological toll of stress, anxiety, and depression. Manual measurement of wound area, meanwhile, is prone to inaccuracy, which can distort clinical assessments of healing progress and delay the detection of complications such as infection.

UlcerNet’s architecture is a dense convolutional encoder-decoder network guided by attention at multiple levels. The input image, resized to 384 by 384 pixels, first passes through a coordinate convolution layer, a clever trick that appends explicit spatial coordinates to the feature map so the network knows where in the image each pixel sits. A folding layer then compresses the spatial dimensions while deepening the channels, producing a 96 by 96 by 80 feature map. From there, the encoder path is built from the paper’s central innovation, the Multi-Scale Residual Attention Dense Block, or MRADB, which fuses multi-scale convolutions, dense connections inspired by DenseNet, channel attention, and residual learning. Dense connectivity means each layer receives the concatenated outputs of all previous layers, enhancing feature reuse and gradient flow, while self-gated residual branches preserve original features and stabilize learning.

Complementing the MRADB blocks is a second novel module, the Modified Receptive Field Block, or MRFB, which sits at the bottleneck of the network to capture global context. The MRFB begins with a 1×1 convolution for dimensionality reduction, then splits into three parallel branches, each combining larger convolutions of 3×3, 5×5, or 7×7 with a 3×3 dilated convolution whose dilation rates of 1, 3, and 5 allow the network to see features at different spatial extents without adding parameters. The branch outputs are concatenated and fused with a 1×1 convolution, and a residual connection with ReLU activation ties the fused representation back to the original input. This design lets UlcerNet gather multi-scale context around ulcers of wildly different sizes, which range in the dataset from 0.04 percent to 35.05 percent of the image area.

Attention is the third pillar of the system. A Convolution Block Attention Module, or CBAM, is applied near the end of the pipeline, sequentially computing channel attention, which highlights salient features by exploiting inter-channel dependencies, and spatial attention, which concentrates on relevant regions by pooling information along the channel axis. The decoder path mirrors the encoder, using deconvolution layers to upsample features and skip connections to reintroduce fine-grained encoder features refined by further MRADB blocks. Three parallel outputs are unfolded to full resolution and fused, then refined by the CBAM before a residual connection combines the attention-refined output with the original unfolded features to produce the final two-class segmentation map. The network is trained with a hybrid loss that weights dice loss, which addresses class imbalance by emphasizing the overlap between predicted and true regions, together with pixel-wise cross-entropy, which penalizes individual misclassifications.

The team trained and evaluated UlcerNet on the public DFUC 2022 dataset, a corpus of 4,000 close-up foot images of which 1,999 annotated images, split into 1,620 training and 379 testing images, were used for supervised learning. Images have a resolution of 640 by 480 pixels, and data augmentation with random flips and rotations added orientation diversity. Training used the Adam optimizer with an initial learning rate of 0.001, reduced by 10 percent every 10 epochs when validation loss stalled, a batch size of 4, and 250 epochs on an NVIDIA GeForce RTX 2070. The results are striking: UlcerNet achieved a Dice coefficient of 73.50 percent and a mean Intersection over Union of 76.04 percent, comfortably ahead of standard baselines such as FCN8 at 49.93 percent, SegNet at 47.68 percent, and DeepLabV3+ at 62.77 percent Dice, and also surpassing recent DFU specialists including TranUNet at 42.61 percent, HarDNet-DFUS at 59.81 percent, and UFOS-Net at 73.09 percent.

An ablation study dissects exactly where the performance comes from. Starting from a plain baseline encoder-decoder with a Dice of 0.6010, adding the coordinate convolution layer, then MRADB, then MRFB, and finally the CBAM attention module raised the score step by step to the full model’s 0.7350 Dice and 0.76047 mIoU. Removing either the coordinate layer or the global attention mechanism caused sharp declines, with the Dice dropping to 0.612 when both components were disabled, confirming that spatial encoding and attention recalibration are essential rather than decorative. The statistical rigor is notable for a deep learning paper: a paired Wilcoxon signed-rank test across the 379 test images showed median paired improvements of +0.0209 in Dice and +0.0290 in IoU over a matched baseline, with p-values below 0.001, and all nine seed-pair combinations remained significant, while three independent training runs produced tightly clustered results of Dice 0.7397 plus or minus 0.0057.

Computational efficiency matters if such models are ever to run in clinics or on mobile devices, and here UlcerNet is impressively lean. The network contains about 7.82 million trainable parameters with a checkpoint size of just 30.27 megabytes, and requires 158.12 GFLOPs per image at 384 by 384 resolution. On an NVIDIA GeForce RTX 4070 SUPER, single-image inference takes 68.59 milliseconds, about 14.6 frames per second, dropping to 34.21 milliseconds per image, or 29.2 frames per second, when images are processed in batches of eight. The authors are careful to state that these measurements demonstrate computational feasibility rather than proven real-time clinical deployment, since practical use also depends on hardware availability, image capture, preprocessing, post-processing, and integration with clinical workflows.

The analysis also probes where the model struggles and why, offering a rare window into the interpretability of medical AI. Per-image Dice scores correlated most strongly with lesion-to-background grayscale contrast at +0.338, while boundary irregularity, quantified as perimeter squared over four pi times area, correlated negatively at -0.248, and lesion area correlated at +0.155, indicating that low-contrast, highly irregular lesions are the hardest to trace. Two typical failure cases illustrate the point: one image with a lesion of 1.82 percent area but contrast of only 2.4 produced severe under-segmentation, and another with a tiny 0.49 percent lesion and an irregularity score of 6.83 also failed completely. Grad-CAM heatmaps and the CBAM attention maps show that in successful cases the network’s attention concentrates tightly on the ulcer and its boundaries, while in failed cases the maps are weaker and more dispersed, reflecting low confidence rather than confidently wrong localization.

The authors acknowledge limitations that temper the headline numbers. UlcerNet was tested only on DFUC 2022 without independent cross-validation, comparisons with newer architectures such as Swin-UNet and MedNeXt were not included, and robustness to variations in illumination, skin tone, camera quality, and other acquisition conditions was not systematically studied. Clinical validation, expert review, and integration into telemedicine platforms were beyond the scope of the work, although the broader context is compelling: telecare systems are already transforming diabetic foot treatment by bringing specialized care to rural settings, and cloud-based deep learning frameworks and platforms such as DFUCare have demonstrated that detection and longitudinal monitoring can be combined into a holistic digital history of healing. With diabetes prevalence rising from 5.5 percent in 1990 to 11.4 percent in India, and about 240 million people worldwide undiagnosed, tools that enable early identification and precise monitoring of foot ulcers could prevent complications before they become life-altering. The team’s next steps include testing UlcerNet on other benchmark datasets, hardening it against real-world imaging variability, and pursuing the expert evaluation needed to move it from the laboratory into the clinic.

Subject of Research: Attention-guided dense convolutional network automates diabetic foot ulcer segmentation

Article Title: An attention-guided dense convolutional network for automated diabetic foot ulcer segmentation

Article References: R. S., R., Nathan, S., & Arun R., A. (2026). An attention-guided dense convolutional network for automated diabetic foot ulcer segmentation. Frontiers in Endocrinology, 17, Article 1940557. https://doi.org/10.3389/fendo.2026.1940557

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1940557

Keywords: attention-guided, dense, convolutional, network, automated, diabetic, foot, ulcer, segmentation, scientific research

The researchers call their architecture ResNet50-MASAG-UNet, and each component of that name encodes a deliberate technical choice. The backbone is a UNet, the encoder-decoder network that has dominated biomedical segmentation since 2015, but with its standard encoder replaced by ResNet50, a 50-layer residual network pretrained on ImageNet. That transfer learning decision is crucial because the team evaluated their system on the DCA1 database, a public collection of just 134 grayscale angiograms of 300 by 300 pixels, each annotated with ground-truth vessel masks prepared by expert cardiologists at the Cardiology Department of the Mexican Social Security Institute. Small datasets like this cause deep encoders trained from scratch to overfit badly; the ablation numbers show it plainly. With random initialization the model reaches a Dice coefficient of 81.2 percent, but with ImageNet-pretrained weights it climbs to 88.3 percent—a gain of more than seven points attributable to pretraining alone.

Before any network sees the images, the pipeline applies a three-stage preprocessing sequence designed to rescue vessel visibility from the raw angiogram’s low signal-to-noise ratio. First, bilateral filtering with a neighbourhood diameter of 9 and color and spatial sigma values of 75 smooths background noise while preserving vessel edges, because bilateral filtering weights neighbours by both spatial distance and intensity similarity. Second, unsharp masking with a 9 by 9 Gaussian kernel, sigma of 10.0, and an amount factor of 1.5 sharpens fine vascular details by subtracting a blurred copy of the image from the original. Third, Contrast Limited Adaptive Histogram Equalization with a clip limit of 3.0 on an 8 by 8 tile grid redistributes local intensity contrast and corrects uneven brightness without amplifying noise in dark regions. The order matters: noise reduction comes first, edge sharpening second, local contrast enhancement last, so that CLAHE does not magnify artefacts the earlier steps have not yet suppressed. The preprocessing ablation is striking—raw images yield 80.2 percent Dice, the full pipeline 88.3 percent, and the boundary error metric HD95 falls from 12.6 pixels to 4.1 pixels.

The genuinely novel element sits at the skip connections, the channels through which a UNet passes encoder features directly to the decoder. In a standard UNet these connections are a liability in angiography: they carry catheter outlines, bone edges, and heart-shadow noise straight into the reconstruction path, weakening the recovery of thin vessels. The team replaced them with a Multi-Scale Adaptive Spatial Attention Gate, or MASAG, adapted from the MSA2Net architecture of Kolahi and colleagues and installed at all four encoder skip levels, with channel dimensions of 64, 64, 128, and 256. Each MASAG operates in four stages. It first extracts multi-scale features from the encoder signal using a depth-wise separable 3 by 3 convolution for fine vessel detail and a dilated convolution with rate 2 for wider structures such as curves and branches, fusing them with global decoder context aligned through global average pooling and upsampling. A Softmax-based spatial selection map then generates two attention weights that reweight the encoder and decoder features, allowing the two paths to compete per pixel for influence.

What follows is the two-way interaction that distinguishes MASAG from earlier attention modules like CBAM or the attention gates of Schlemper and colleagues. The reweighted encoder and decoder features cross-modulate each other—the encoder signal modulates the decoder guidance and vice versa—producing an interaction feature that captures joint vessel-relevant information from both directions. Finally, a channel attention vector computed with a reduction ratio of 8 recalibrates the output before it reaches the decoder. The ablation table makes the case quantitatively: within the identical ResNet50-UNet backbone, an attention gate achieves 85.4 percent Dice, CBAM 86.6 percent, and MASAG 88.3 percent. Removing MASAG entirely drops performance by roughly 8 to 10 Dice points, and equipping all four skip connections outperforms one, two, or three, with the four-skip configuration reaching the full 88.3 percent. Image-level paired statistical tests across all 134 out-of-fold predictions confirm the gains are significant, with Cohen’s d effect sizes of 1.55 or larger and p-values below 0.001 after Bonferroni correction.

Attention alone, however, cannot guarantee that the predicted vessel tree remains anatomically connected. Binary cross-entropy and Dice losses optimize pixel overlap and region similarity, but a prediction can score well on both while fragmenting the centreline into disconnected pieces—a fatal flaw for any downstream stenosis analysis that depends on tracing a continuous path along an artery. The team therefore trained with a hybrid loss combining three terms in equal weights: binary cross-entropy for pixel-level refinement, Dice for region overlap and class imbalance, and centreline Dice, or clDice, for topological connectivity. clDice, introduced by Shit and colleagues in 2021, compares the centreline structures of prediction and ground truth using a differentiable soft-skeletonization operation in which iterative min and max pooling with a 3 by 3 kernel approximate morphological erosion and dilation. Five iterations suffice to extract thin coronary skeletons at the 256 by 256 working resolution. When a vessel breaks, the skeleton fragments, both centreline-matching terms fall, and the loss rises—forcing the network to learn connectivity as a training objective rather than hoping for it as a by-product.

The topology-aware loss earns its place in the numbers. Adding clDice to a Dice-plus-BCE baseline lifts Dice from 85.9 to 88.3 percent, the clDice metric itself from 86.3 to 89.1 percent, HD95 from 7.4 to 4.1 pixels, and the Betti-0 error—a count of connected components in the prediction versus the ground truth, using 8-connectivity appropriate for thin diagonally oriented vessels—from 3.1 to 1.5. A sensitivity analysis varying the clDice loss weight from 0.1 to 5.0 shows performance stable across the range, peaking at the balanced equal-weight setting, so the choice is not a fragile hyperparameter. The team also reports the 95th percentile Hausdorff distance to capture boundary-level accuracy and penalize large local deviations, and they candidly note a limitation of Betti-0 itself: because it compares only component counts, a false merge and a false split can offset each other and leave the count unchanged, so they interpret it alongside overlap and boundary metrics rather than in isolation.

Training followed a two-phase schedule on an NVIDIA Tesla P100 GPU in the Kaggle environment using Keras-TensorFlow with Python 3.10. For the first 25 epochs the ResNet50 encoder stayed frozen while the MASAG gates and decoder learned stable features at a learning rate of 1 times 10 to the minus 4; the encoder was then fully unfrozen for another 25 epochs of joint fine-tuning, a strategy the ablation confirms outperforms both a permanently frozen encoder at 84.1 percent Dice and selective unfreezing of deeper stages at 86.5 percent. The complete model carries 37,397,633 parameters, requires roughly 32.5 GFLOPs per forward pass at 256 by 256 input, trains in about four minutes per fold, and infers in approximately 22 milliseconds per image—fast enough, in principle, for clinical workflows. Five-fold cross-validation, with four folds splitting 107 training and 27 validation images and one fold 108 and 26, produced the reported means and standard deviations, and the learning curves show training and validation losses declining in close alignment from roughly 0.90 to 0.01 and 0.88 to 0.02 respectively, with narrow shaded bands across folds indicating stable generalization without major overfitting.

The headline comparison is blunt: the standard UNet baseline reaches 74.0 percent Dice, UNet++ 74.7 percent, and a ResNet50-UNet without attention 76.1 percent, while the full proposed model attains 88.3 percent Dice, 79.1 percent Jaccard, 86.4 percent precision, and 90.5 percent recall. Against recently reported methods on DCA1—including FiTNet at 81.1 percent, EAG-SAE-ESTF at 80.6 percent, and CoorMa at 80.4 percent—the proposed model reports among the highest Dice, Jaccard, and recall values, though the authors are careful to state that these studies used different preprocessing and evaluation protocols, so the comparison is a reference rather than a strictly controlled head-to-head benchmark. Qualitative results, with true positives in green, false positives in red, and false negatives in blue, show the model maintaining a consistent balance between error types and suppressing non-vascular activations from catheters and bones without any separate removal step.

The authors are equally candid about the boundaries of the achievement. The evaluation rests on a single public dataset of 134 images with cross-validation folds rather than a separate test set, so performance may appear slightly better than it actually is, and patient-level independence between training and validation splits cannot be verified because the distributed images carry no patient identifiers. Thin distal vessel branches are sometimes missed, very low-contrast segments can blend into surrounding tissue, and unsharp masking and CLAHE may amplify background noise in some areas despite bilateral filtering and a moderate clip limit. Most tellingly, a zero-shot test on the external XCAD dataset of 126 images from a different imaging system—resized but not preprocessed—yielded a Dice of just 22.3 percent, a sobering measure of the domain gap between imaging systems and a reminder that cross-dataset adaptation remains future work. No clinical tasks such as stenosis quantification were evaluated, and the authors note that expert-annotated stenosis data, which DCA1 does not provide, would be required for blockage percentage analysis. Still, the structural basis is now in place: a segmentation that keeps the coronary tree connected, suppresses the visual clutter of the catheter lab, and runs in milliseconds. Whether that foundation survives contact with diverse hospital data from different scanners is the question the next phase of this research must answer.

Subject of Research: Artificial Intelligence

Article Title: Topology-preserving coronary vessel segmentation in X-ray angiography using ResNet50-MASAG-UNet

Article References: Sunilkumar, G., & Kumaresan, P. (2026). Topology-preserving coronary vessel segmentation in X-ray angiography using ResNet50-MASAG-UNet. Frontiers in Radiology, 6, Article 1949450. https://doi.org/10.3389/fradi.2026.1949450

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1949450

Keywords: coronary artery disease, X-ray coronary angiography, deep learning, vessel segmentation, attention gates, topology-preserving loss, ResNet50, medical imaging, Topology-preserving, coronary, vessel, segmentation

The system, called PDSA-DBUNet, was developed by Jiatong Shang of Shangluo University, Guolin He of Heping Hospital Affiliated to Changzhi Medical College, and Xuyang Li of Shaanxi University of Science and Technology. It is a dual-branch U-Net that runs a convolutional neural network and a Transformer in parallel, letting each branch do what it does best. The convolutional branch preserves local textures, edges, and boundary details, while the Transformer branch models global semantic dependencies across the whole image. The innovation lies in how the Transformer branch is made dramatically cheaper: instead of computing attention over every pixel token, it compresses the image into a small set of informative sparse tokens and then mixes them with a polynomial operation rather than conventional multi-head self-attention.

To understand why this matters, consider the computational anatomy of standard Transformers. Self-attention computes pairwise relationships among all tokens, so its cost grows quadratically with the number of tokens. On a high-resolution CT slice, most tokens represent background tissue that carries little diagnostic value, yet dense attention still spends enormous computation correlating them with one another. The researchers’ dynamic sparse visual token generation module addresses this by using learnable spatial importance weights, produced through a global projection and a Softmax operation, to adaptively aggregate dense tokens into a compact set of discriminative sparse tokens. An element-wise modulation step, normalized by a Sigmoid function, further refines each sparse token so that the aggregation responds to the actual content of the image. Organ regions, tumor areas, and blurred boundaries are emphasized, while irrelevant background responses are suppressed.

Crucially, the mapping between sparse tokens and their original dense spatial positions is retained. After the Transformer branch processes the compact token set, the features are projected back to their original spatial locations and reshaped into feature maps that can be fused with the convolutional branch at matching resolutions. This restoration step is what makes the sparse design compatible with pixel-level segmentation, a requirement that many general-purpose efficient Transformer methods, such as Perceiver or DynamicViT, were never designed to satisfy for dense medical prediction.

The second efficiency leap comes from replacing self-attention entirely with a Polynomial Mixer, based on the PoM mechanism. Rather than building an explicit attention matrix, the mixer aggregates all sparse tokens into a shared global state using learnable polynomial coefficients and element-wise power operations, then applies a Sigmoid-based gating mechanism so each token can selectively retrieve relevant contextual information from that shared state. The computational complexity of this operation scales linearly with the number of tokens rather than quadratically. According to the authors’ complexity analysis, the overall cost of the PDSA module is dominated by sparse token generation and polynomial mixing, both of which are far cheaper than dense attention when the sparse token count is much smaller than the original token count, and no dense attention matrix ever needs to be stored in memory.

Bridging the two branches is the third key component: a Dual-Scale Parallel Attention Fusion module, or DSPAF. The researchers argue that simply adding or concatenating convolutional and Transformer features is inadequate because the two feature spaces respond inconsistently to the same anatomy. DSPAF instead operates at two complementary scales in parallel. A global channel attention path uses global average pooling and fully connected layers to model semantic relationships between the branches, providing global contextual guidance to the CNN features and local structural constraints to the Transformer features. A local spatial interaction path concatenates the aligned features and applies convolutional processing to highlight locations jointly activated by both branches, refining boundaries, textures, and small target regions. A residual connection preserves original local details and stabilizes training.

The experimental evidence is striking. On the LiTS liver tumor segmentation benchmark, using 131 contrast-enhanced abdominal CT cases split into 100 training and 31 testing cases, PDSA-DBUNet achieved a Dice score of 96.79 percent for liver segmentation, exceeding Swin UNETR by 1.66 percentage points and nnU-Net by 2.77 percentage points. For liver tumors, which are small, irregular, and low in contrast, it reached 86.13 percent Dice, outperforming Swin UNETR and nnU-Net by 1.22 and 1.26 percentage points respectively, and beating the conventional U-Net baseline by nearly four points. On the tumor task it also achieved the best results on all four error-based metrics, including a volumetric error of just 0.76 percent and boundary distances substantially lower than every competing method.

The KiTS19 kidney dataset told a similar story. Across 210 abdominal CT volumes divided into 168 training, 21 validation, and 21 testing cases, the model scored 94.03 percent Dice for kidney segmentation and 89.06 percent for kidney tumors, leading on all three reported metrics for the tumor task, including an average symmetric surface distance of 1.91 millimeters and a maximum surface distance of 17.09 millimeters. Qualitative comparisons showed the network producing more complete organ masks and more compact tumor contours with fewer false positives and false negatives than U-Net, TransUNet, Swin-Unet, nnU-Net, Swin UNETR, and other baselines, particularly near low-contrast tissue interfaces where rival methods leaked or under-segmented.

Perhaps the most persuasive numbers come from the ablation study. A plain U-Net baseline reached 93.99 percent Dice with 65.47 GFLOPs and 34.53 million parameters. Simply bolting on Transformer blocks, whether stacked or parallel, raised accuracy by only fractions of a point while inflating the model to as much as 496.45 million parameters and 673.56 megabytes of storage. Introducing the PDSA module into the parallel structure lifted Dice to 96.14 percent while cutting FLOPs to 70.04 G and parameters to 40.42 million. Adding DSPAF brought the final score to 96.79 percent, with the complete model requiring 72.87 GFLOPs, 41.62 million parameters, and a checkpoint of just 129.93 megabytes. Compared with the parallel CNN-Transformer variant without sparse attention, the final architecture improved Dice by 2.68 percentage points while reducing computation by 50.19 percent, parameters by 91.62 percent, and model size by 80.71 percent.

The authors are candid about limitations. The current experiments operate on 2D CT slices, which do not exploit inter-slice continuity in volumetric scans, and validation so far rests on two public datasets rather than multi-center clinical data. Future work, they write, will extend the architecture to three-dimensional or 2.5-dimensional inputs and compare directly with 3D variants of nnU-Net, nnFormer, and Swin UNETR under identical settings. Even so, the core message is likely to resonate across medical AI: the quadratic cost of dense attention is not a price that must be paid for global context. By deciding dynamically which image regions deserve to be tokens at all, and by mixing those tokens with polynomial algebra instead of pairwise attention, PDSA-DBUNet demonstrates that a network can see the whole abdomen clearly while thinking about far less of it at a time.

Subject of Research: A dual-branch CNN-Transformer neural network with polynomial dynamic sparse attention for abdominal CT organ and tumor segmentation

Article Title: PDSA-DBUNet: dual-branch U-Net for abdominal medical image segmentation driven by polynomial dynamic sparse attention

Article References: Shang, J., He, G., & Li, X. (2026). PDSA-DBUNet: dual-branch U-Net for abdominal medical image segmentation driven by polynomial dynamic sparse attention. Frontiers in Signal Processing, 6, Article 1904505. https://doi.org/10.3389/frsip.2026.1904505

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1904505

Keywords: medical image segmentation, abdominal CT, deep learning, U-Net, Transformer, sparse attention, polynomial mixer, CNN-Transformer hybrid, liver tumor segmentation, kidney segmentation, LiTS dataset, KiTS19

The clinical problem the researchers set out to solve is one that radiologists know all too well. Breast lesions classified as BI-RADS 4 or 5 — meaning suspicious enough to warrant biopsy — cover a wide spectrum of malignancy risk, and their ultrasound appearances overlap maddeningly. A benign lesion can look irregular in shape with indistinct margins and a dark, hypoechoic texture; so can a cancer. Posterior acoustic attenuation, a shadowing effect once considered a red flag, appears in both. This morphological ambiguity produces high inter-observer variability, particularly among less experienced practitioners. Ultrasound diagnosis is also highly operator-dependent, shaped by scanning skill, image acquisition standards, and subjective judgment. The stakes are high on both sides: a missed malignancy delays treatment, while an unnecessary core-needle biopsy or surgery exposes a patient to risk, anxiety, and cost for what turns out to be a harmless finding.

The challenge is compounded by breast tissue density. Asian women, on average, have denser breast tissue than other populations, which degrades the performance of mammography, the traditional screening mainstay. In China, clinical guidelines therefore pair mammography with ultrasound as the primary and auxiliary imaging methods for breast cancer detection. Ultrasound is convenient, radiation-free, and particularly effective in dense tissue — but its accuracy in distinguishing benign from malignant BI-RADS 4–5 lesions has remained stubbornly limited. Previous computational efforts largely focused on static images, requiring the sonographer or an algorithm to pick a single representative frame from a live scan. That choice discards a wealth of information: the way a lesion’s boundaries shift as the probe moves, how its internal echoes evolve, and the temporal context that an experienced human eye implicitly absorbs during a real-time examination.

TGMIL was designed specifically to exploit that discarded information. The architecture treats each ultrasound video as a bag of frames in a machine learning framework known as multi-instance learning, where the overall diagnosis is known but the diagnostic value of each individual frame is not. Each frame is first encoded by a VGG-13 convolutional backbone, generating a sequence of frame-level features. Crucially, the model also encodes the temporal order of frames through a linear layer, concatenating this positional information with the visual features so the network can learn from how a lesion appears, evolves, and disappears across the scan. An attention module then adaptively assigns higher weights to the most diagnostically informative frames and aggregates them into a single video-level representation. In effect, TGMIL learns which moments of the scan matter — suppressing redundant frames and emphasizing clinically relevant temporal segments — rather than treating every image equally.

The study enrolled 532 breast lesions categorized as BI-RADS 4–5 across four hospitals between November 2019 and January 2023. Of these, 334 lesions (61 benign and 273 malignant) formed the training and validation sets, while 198 lesions (50 benign and 148 malignant) were reserved as an independent test set. All videos were acquired with high-frequency linear probes, scanned continuously from the first appearance of each lesion to its complete disappearance, and converted to DICOM format. Preprocessing was rigorous: videos were converted to a lossless format, black borders and text overlays were stripped, a K-means algorithm extracted 15 key frames per video, and pixel values were standardized. Every diagnosis was confirmed by pathology from surgery or ultrasound-guided biopsy, and all examinations from the same patient were kept within a single dataset partition to prevent data leakage.

On the external test set, TGMIL achieved an area under the receiver operating characteristic curve of 0.824, decisively outperforming mainstream video classification architectures including 3D-CNN (AUC 0.665), CSN (0.613), and R(2+1)D-50 (0.678). It also beat static image-based classifiers: ResNet50, VGG16, the transformer-based ViT B/32 (which managed only 0.656), and even BU-Mamba, a recently proposed static ultrasound model. BU-Mamba achieved remarkable sensitivity of 0.988 but suffered from low precision of 0.752, indicating a tendency toward over-diagnosis. TGMIL struck a better balance, pairing high sensitivity of 0.966 with precision of 0.836 — meaning it reliably caught malignancies without flooding the clinic with false alarms. Ablation tests confirmed the value of the attention mechanism: simple mean pooling and max pooling baselines reached AUCs of only 0.752 and 0.770 respectively, showing that adaptively identifying informative frames, rather than averaging or maximizing across them, is what drives the performance gain.

The most clinically compelling results came from the reader study. Nine ultrasonographers — three juniors with under three years of experience, three attendings with four to nine years, and three seniors with a decade or more — independently reviewed the 198 test lesions, first without AI assistance and then, four weeks later and in a reshuffled case order, with TGMIL’s real-time benign-versus-malignant probability estimates. Unaided, the model itself achieved a patient-based AUC of 0.845, significantly outperforming both junior sonographers (0.682) and attendings (0.802). With TGMIL’s help, junior accuracy jumped from 66.7 to 89.2 percent and sensitivity surged from 67.4 to 94.2 percent, while attending accuracy rose from 80.8 to 90.0 percent. Across all readers, pooled accuracy improved from 77.5 to 95.0 percent, and the pooled AUC climbed from 0.871 to 0.964. Specificity for the whole group rose from 88.2 to 97.1 percent — a direct strike against unnecessary biopsies.

Interestingly, the most experienced readers gained the least. Senior ultrasonographers, who already achieved an AUC of 0.903 unaided, showed no statistically significant improvement on any metric with AI assistance, and their AUC dipped slightly from 0.898 to 0.885. The authors offer a candid explanation: because TGMIL’s standalone performance (0.824) falls below that of seasoned experts, the model may add little information beyond what veterans already extract from morphology, Doppler signals, and clinical intuition. Worse, when model predictions conflict with expert judgment, uncertain outputs can introduce cognitive interference or automation bias, occasionally nudging a correct decision astray. The study’s failure analysis reinforced this caution: some benign lesions displayed pseudo-malignant features such as spiculated margins that fooled the model into false positives, while some cancers with weak or atypical vascular signals escaped detection as false negatives.

The practical implications are nonetheless considerable. TGMIL was trained on a single NVIDIA RTX 3090 GPU — no supercomputer required — and once trained, inference on ultrasound video is computationally light enough for a local workstation, meaning the system could be integrated directly into the ultrasound workflow. The model also showed strong probability calibration (Brier score 0.071) and robust performance across all four external test centers, with AUCs ranging up to 0.886, suggesting it generalizes across different scanners, hospitals, and patient populations. For regions facing shortages of senior breast imaging specialists, an AI assistant that lifts junior sonographers to near-expert accuracy could meaningfully expand access to reliable diagnosis.

Limitations remain. The cohort was hospital-based and highly selected: only pathologically confirmed BI-RADS 4–5 lesions larger than 5 millimeters were included, excluding prostheses, poor-quality videos, and other BI-RADS categories. No asymptomatic screening population was tested, and screening settings — with lower disease prevalence and smaller incidental lesions — may behave very differently. The authors themselves urge prospective validation in screening cohorts before broader deployment. Still, the study marks a conceptual shift in medical AI: from judging isolated snapshots to understanding the full temporal narrative of a scan, the way a human expert does. If subsequent trials confirm these findings, the humble ultrasound video — long a passive record of a scan — may become an active diagnostic instrument, and the gap between novice and veteran sonographers may finally begin to close.

Subject of Research: Deep learning analysis of ultrasound videos for classifying benign and malignant BI-RADS 4–5 breast lesions

Article Title: AI-enhanced ultrasound video analysis for BI-RADS 4–5 breast lesions: a multicenter study of diagnostic performance

Article References: Li, W.-B., Du, Z.-C., Huang, W.-H., Qiu, R.-Y., Chen, Y.-C., Lu, L.-S., Ji, Z.-Y., Guo, L.-H., He, M.-Q., Wang, L.-S., & Zhang, G.-J. (2026). AI-enhanced ultrasound video analysis for BI-RADS 4–5 breast lesions: a multicenter study of diagnostic performance. Holistic Integrative Oncology, 5(1), Article 80. https://doi.org/10.1007/s44178-026-00302-8

Image Credits: AI Generated

DOI: 10.1007/s44178-026-00302-8

Keywords: artificial intelligence, deep learning, breast cancer, ultrasound, BI-RADS, computer-aided diagnosis, multi-instance learning, medical imaging, diagnostic accuracy, oncology, radiology, multicenter study

The study, led by Qiang Ma and TianXi Liang with corresponding author Yin Wang of Shanghai Pulmonary Hospital, was designed as a retrospective, multicenter analysis. The team assembled ultrasound data from 1,059 patients with subpleural pulmonary lesions, drawn from a primary center between January 2023 and December 2024, with additional external cohorts recruited from two other hospitals between June and December 2024. The patient population had a mean age of roughly 61 years and was predominantly male, with 726 male patients among the 1,059 enrolled. Crucially, the researchers did not simply test their models on data from the same hospital that produced the training set. They held out two entirely separate external test sets, comprising 108 and 94 patients respectively, to see whether the artificial intelligence could generalize to machines, protocols, and patient populations it had never encountered.

The architecture of the framework reflects a logical division of labor. The first stage is segmentation: the computer must find the lesion in the ultrasound image and trace its boundaries precisely. To do this, the team benchmarked three well-established convolutional neural network approaches: nnU-Net, U-Net, and DeepLabv3+. The nnU-Net, a self-configuring variant of the U-Net family that has become something of a gold standard in medical image segmentation, emerged as the clear winner. Its performance was measured with the Dice similarity coefficient, a metric that scores the overlap between the machine-drawn region of interest and the ground-truth annotation, where a value of 1.0 indicates perfect agreement. The nnU-Net achieved a Dice score of 0.920 on the internal tuning-validation set of 257 cases, and, more importantly, held its ground externally, scoring 0.913 on External Test Set 1 and 0.915 on External Test Set 2. Those numbers mean the model’s outlines agreed with expert annotations to within a few percent of perfect overlap, even on unfamiliar data.

With the lesion reliably isolated, the second stage took over: classification. Here the researchers cast a wide net, training and comparing thirteen different candidate architectures drawn from the most influential families of convolutional networks in modern computer vision. The lineup included DenseNet121 and DenseNet201, the EfficientNet family from B0 through B5, three depths of ResNet (18, 50, and 101), Inception-V3, and VGG19. Each architecture brings a different inductive bias to the task. DenseNets, for instance, connect every layer to every subsequent layer, encouraging feature reuse and parameter efficiency; EfficientNets scale depth, width, and image resolution in a carefully balanced way; ResNets use skip connections that allow very deep networks to train without vanishing gradients. After head-to-head evaluation on the internal tuning-validation set, DenseNet121 proved the strongest performer and was selected as the final classification model.

The headline result is the model’s ability to distinguish benign from malignant lesions on images it had never seen. On the two external cohorts, DenseNet121 achieved areas under the receiver operating characteristic curve of 0.889 (95% confidence interval 0.827 to 0.951) and 0.865 (95% confidence interval 0.790 to 0.940). The AUC is a threshold-independent measure of discriminative ability: a value of 0.5 would be no better than a coin flip, while 1.0 would be perfect. Values approaching 0.9 indicate that, across the full range of possible decision thresholds, the model ranks a randomly chosen malignant lesion as more suspicious than a randomly chosen benign one roughly nine times out of ten. For a task that even experienced human readers find difficult on grayscale ultrasound alone, that level of performance is striking, and the fact that it held up across two independent hospitals is what separates this work from many single-center proof-of-concept studies.

Perhaps the most clinically consequential part of the study, however, was not the model’s raw scores but a carefully designed reader study. The researchers recruited six radiologists, three junior and three senior, and compared their diagnostic accuracy under two conditions: reading unaided, and reading with the AI’s output available as an aid. Because each radiologist read the same cases under both conditions, the analysis had to account for the fact that multiple observations come from the same readers and the same patients. The team used case-cluster bootstrap methods to estimate pooled three-reader differences, applying Bonferroni adjustment to control for multiple statistical comparisons, a conservative approach that reduces the risk of false-positive findings.

The outcome was unambiguous for the less experienced readers. Diagnostic accuracy among junior radiologists rose significantly in the AI-assisted reading session in both external test cohorts, with Bonferroni-adjusted P values below 0.001 in both cases. In practical terms, the AI functioned as a leveler of expertise: a junior radiologist armed with the model’s segmentation and classification output approached the kind of judgment that would otherwise require years of accumulated pattern recognition. For senior radiologists, the improvements, while directionally positive, did not reach statistical significance after the multiplicity adjustment. That pattern is common in AI-assisted reading studies and makes intuitive sense. Experts already perform well on their own, leaving less headroom for an algorithm to add measurable value, whereas novices benefit most from a second opinion that never tires and never varies.

The technical choices behind the framework deserve attention because they address two chronic weaknesses of medical deep learning: overfitting to a single center and the black-box problem. By validating on two external cohorts, the authors demonstrated that their pipeline was not merely memorizing the imaging characteristics of one hospital’s ultrasound machines. By comparing thirteen architectures rather than anointing a single favorite in advance, they let the data decide which inductive biases best suited the texture and morphology cues that distinguish benign inflammation from malignancy in subpleural tissue. The study also references gradient-weighted class activation mapping, or Grad-CAM, a visualization technique that highlights which regions of an image drove the model’s decision, and decision curve analysis, which evaluates the net clinical benefit of a diagnostic tool across different threshold preferences. The inclusion of these tools signals an effort to make the framework not only accurate but interpretable and clinically actionable.

The stakes of this problem are easy to underestimate. Subpleural lesions sit at a diagnostic crossroads: they can be early lung cancers, which demand prompt resection or oncological treatment, or they can be benign infectious or inflammatory processes, which demand nothing more invasive than follow-up or antibiotics. Distinguishing between them on grayscale ultrasound alone is genuinely hard, because both categories can produce hypoechoic masses with irregular margins, and the acoustic features that radiologists rely on are subtle and operator-dependent. An error in either direction carries a cost: a missed malignancy forfeits the best window for curative treatment, while an unnecessary biopsy or resection of a benign lesion exposes the patient to procedural risk and expense. A reliable computational second opinion, delivered in seconds from an image that is already being acquired, could shift that balance meaningfully.

The authors are appropriately measured in their conclusions, describing the framework as a potential auxiliary diagnostic tool rather than a replacement for clinical judgment. The study is retrospective, and its models were trained and tested on de-identified data from three Chinese hospitals, so prospective validation in other health systems, with different scanners and patient demographics, remains a necessary next step. The work was supported by the National Natural Science Foundation of China and several Shanghai municipal and institutional research funds, and it was conducted under ethics approvals from Shanghai Pulmonary Hospital, Shanghai Chest Hospital, and Xi’an Chest Hospital in compliance with the Declaration of Helsinki and STARD reporting guidelines. Still, the combination of near-expert segmentation, strong external AUCs, and a demonstrable boost to junior clinicians’ accuracy makes a compelling case that the humble grayscale ultrasound image, long dismissed as too subjective for lung lesion triage, is becoming a far more powerful instrument in the hands of a well-trained neural network. If future prospective trials confirm these findings, the technology could bring expert-level lung lesion assessment to clinics where CT scanners are scarce and thoracic radiologists are scarcer, turning a bedside probe and a laptop into a credible first line of defense against one of the world’s deadliest cancers.

Subject of Research: Deep learning for ultrasound-based segmentation and benign-malignant classification of subpleural pulmonary lesions

Article Title: A grayscale ultrasound-based two-stage deep learning framework for automatic segmentation and benign-malignant differentiation of subpleural pulmonary lesions

Article References: Ma, Q., Liang, T., Yi, J., Bai, H., Li, Y., Shen, M., Bi, K., & Wang, Y. (2026). A grayscale ultrasound-based two-stage deep learning framework for automatic segmentation and benign-malignant differentiation of subpleural pulmonary lesions. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02887-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02887-y

Keywords: lung cancer, deep learning, ultrasound, subpleural pulmonary lesions, image segmentation, nnU-Net, DenseNet121, benign-malignant classification, computer-aided diagnosis, radiology, medical imaging, artificial intelligence

The research team, led by Zhe Zhao of the Fourth Medical Center of Chinese PLA General Hospital together with collaborators at the Institute of High Energy Physics of the Chinese Academy of Sciences, the University of Chinese Academy of Sciences, and Minzu University of China, set out to solve two intertwined problems. The first is technical: even state-of-the-art deep learning systems struggle to segment thin, curvilinear structures like cartilage, where a single-pixel error at the boundary can meaningfully distort thickness measurements. The second is infrastructural: high-quality, publicly available annotated datasets for knee MRI are scarce, and datasets reflecting Asian populations, whose anatomical characteristics and disease patterns may differ from the Western cohorts that dominate the literature, are scarcer still. Without representative training data, the authors argue, clinical translation of automated segmentation tools stalls before it begins.

To address the data gap, the researchers constructed what they call the OAMRI dataset, a comprehensively annotated collection of 893 T2-weighted sagittal knee MRI images drawn from 47 Chinese patients. The annotations were produced with the assistance of orthopedic experts and cover four anatomical structures that together define the load-bearing architecture of the knee joint: the femur, the femoral cartilage, the tibia, and the tibial cartilage. T2-weighted sagittal imaging is a mainstay of musculoskeletal radiology because it highlights both morphology and fluid-sensitive changes associated with cartilage degeneration, making it a natural choice for a dataset intended to support early diagnosis. By releasing this resource openly, the team has created one of the few expert-annotated knee MRI datasets rooted in an Asian patient population, a contribution that could meaningfully improve the demographic generalizability of future models.

On the algorithmic side, the researchers chose Swin-Unet as their baseline architecture, a pure Transformer network designed for medical image segmentation. Unlike conventional convolutional neural networks, which process images through local filters and build up context layer by layer, Vision Transformers divide the image into patches and use self-attention mechanisms to weigh relationships between distant regions of the scan. Swin-Unet refines this idea with a hierarchical, U-shaped encoder-decoder design and a shifted-window attention scheme, in which self-attention is computed within small local windows that are then shifted between successive layers. This windowed approach keeps the computational cost manageable while still allowing information to flow across the entire image, a property that matters for cartilage because the correct boundary of a thin structure often depends on context far beyond the local pixel neighborhood.

The team did not simply deploy the architecture off the shelf. They tailored and optimized the Swin-Unet baseline specifically for knee structures, evaluating the cross-entropy loss function that guides training and employing the Adam optimizer, an adaptive moment estimation method that adjusts learning rates parameter by parameter. Cross-entropy loss, which penalizes the discrepancy between predicted and true class labels pixel by pixel, was assessed for its suitability to the multi-class problem of simultaneously distinguishing bone from two distinct cartilage compartments. These seemingly modest engineering choices, the authors report, proved decisive in squeezing reliable performance out of the network on thin anatomical targets where generic configurations often falter.

The evaluation was designed with unusual rigor for this field. In an initial, exploratory slice-level split, where individual MRI slices from the same patient could appear in both training and test sets, the optimized model achieved a theoretical upper-bound Dice similarity coefficient of 95.45 percent for femoral cartilage and 95.84 percent for tibial cartilage. The Dice coefficient, the standard overlap metric in segmentation research, ranges from 0 to 1 and expresses how well the predicted segmentation matches the expert annotation; scores above 90 percent are generally considered excellent. But slice-level splits are known to inflate results, because adjacent slices from the same knee are nearly identical and effectively leak information into the test set. Recognizing this, the researchers also evaluated the model under a strict patient-level split, in which entire patients were held out from training, providing a far more honest test of how the system would behave on genuinely unseen individuals.

Even under that demanding patient-level evaluation, the model held up impressively. It achieved a Dice score of 90.62 percent for femoral cartilage and 89.85 percent for tibial cartilage, averaging 91.43 percent across all four segmented structures. The gap between the slice-level and patient-level results, roughly four to six percentage points, quantifies how much easier the task becomes when the model has already glimpsed the same knee, and the fact that performance remained above 90 percent for both cartilage compartments without that crutch is a strong signal of true clinical generalization. The team further validated the approach on the public MICCAI SKI10 benchmark, a widely used knee MRI segmentation challenge dataset, where the model scored 90.36 percent and 93.98 percent Dice, outperforming state-of-the-art comparison methods on both the proprietary OAMRI data and the public benchmark.

Why does this matter beyond the leaderboard? Quantitative cartilage segmentation is the foundation of modern osteoarthritis assessment. Standard clinical scoring systems such as the Kellgren-Lawrence grading and the Whole-Organ Magnetic Resonance Imaging Score rely on radiologist judgment, which is valuable but coarse and variable. Automated, pixel-accurate maps of femoral and tibial cartilage open the door to precise measurements of cartilage thickness, volume, and surface area, and to tracking subtle progression over months or years, long before joint replacement becomes the only option. The authors point specifically to two clinical applications: quantitative assessment of osteoarthritis severity, and decision support for cartilage repair and replacement therapies, where surgeons need exact knowledge of what cartilage remains and where it has worn away.

The study also carries a quieter but equally important message about how medical artificial intelligence should be built. The pipeline, from ethics approval obtained under the Declaration of Helsinki at the Fourth Medical Center of Chinese PLA General Hospital, through expert annotation of a retrospective, anonymized image cohort, to open release of the dataset under a Creative Commons license, models a transparent path from clinic to algorithm. The work was supported by China’s National Key Research and Development Program, the National Natural Science Foundation of China, and programs of the Chinese Academy of Sciences, reflecting sustained national investment in computational medical imaging. The open-access publication means that research groups anywhere can download the OAMRI annotations, benchmark their own models against the reported results, and extend the resource with additional patients and imaging sequences.

Limitations remain, as the authors would be the first to acknowledge. The dataset, while carefully annotated, comprises 47 patients, and scaling to the diversity of scanners, field strengths, and disease stages encountered in routine radiology will require larger and more heterogeneous cohorts. The model was validated on T2-weighted sagittal images, and performance on other sequences or planes has not been established here. Still, the combination of a patient-level validation protocol, cross-dataset confirmation on SKI10, and an openly shared expert-annotated resource marks this study as a template for clinically meaningful segmentation research. If automated cartilage mapping is to move from the laboratory into the radiology reading room, it will need exactly this kind of foundation: honest evaluation, representative data, and models that keep their accuracy when they meet a patient the algorithm has never seen before.

Subject of Research: Deep learning segmentation of knee cartilage on MRI using an expert-annotated clinical dataset

Article Title: Knee cartilage segmentation on MRI using an expert-annotated clinical dataset

Article References: Zhao, Z., Xu, D., Xu, C., Xi, Z., Huang, H., Pu, Q., & Zhao, L. (2026). Knee cartilage segmentation on MRI using an expert-annotated clinical dataset. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02719-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02719-z

Keywords: knee osteoarthritis, cartilage segmentation, MRI, Swin-Unet, deep learning, OAMRI dataset, Transformer, medical imaging, SKI10 benchmark, Dice similarity coefficient, patient-level validation, clinical decision support

The study, authored by Yang Li and Hongwei Deng of Hengyang Normal University in China, introduces MAF-Net, a multi-scale attention fusion network designed specifically for segmenting the pancreas from abdominal CT images. The work was published on 3 October 2026 in Applied Intelligence, a peer-reviewed journal covering artificial intelligence research. Rather than proposing an entirely new family of neural networks, the authors focused on refining two of the most persistent weaknesses in existing segmentation pipelines: the inability of standard convolutional layers to capture anatomical structures that appear at very different scales, and the difficulty of deciding how much weight to give fine local detail versus broad global context when the two sources of information disagree.

The first of these weaknesses is addressed by what the authors call the Multi-scale Fusion with Spatial Attention block, or MFSA block. In conventional convolutional neural networks, each layer applies filters of a fixed size, which means the network samples the image through a window of one particular spatial extent. Organs like the pancreas, however, do not respect such uniformity. In one slice of a CT volume the organ may appear as a thin, elongated sliver; in another it may swell into a bulbous head region several times wider. A single filter size struggles to describe both. The MFSA block solves this by running parallel convolution branches with different kernel sizes simultaneously, so that contextual information is extracted at multiple scales at once. The outputs of these branches are then combined, giving the network a richer description of the local anatomy than any single scale could provide.

Crucially, the multi-scale convolutions are paired with a spatial attention mechanism. Spatial attention, in the vocabulary of modern computer vision, is a learned weighting map that tells the network which pixel locations in a feature map deserve emphasis and which can be downplayed. For pancreas segmentation this is more than a fashionable add-on. Because the organ occupies a small fraction of each abdominal slice, most of the image is background that could, in principle, distract the model. By learning to boost activation in the pancreatic region and suppress it elsewhere, the attention mechanism sharpens the network’s structural awareness and improves its ability to localize the organ’s often ambiguous boundaries. The authors report that this combination of multi-scale context extraction and spatially guided emphasis is what allows MAF-Net to hold onto the pancreas’s fine contours even in slices where the organ nearly dissolves into its surroundings.

The second core component of the architecture is the Attention Fusion Module, or AFM, which tackles a different question: when a network has both local detail features and global semantic features, how should it blend them? Early layers of a segmentation network capture textures, edges and small-scale patterns, while deeper layers encode the overall layout of the abdomen and the rough location of major organs. A naive fusion, such as simply adding or concatenating these feature sets, treats every location and every channel equally, which can wash out the very details that matter most at organ boundaries. The AFM instead dynamically allocates weights between local details and global semantics, adjusting the balance according to what the image content requires. In regions where the pancreas boundary is crisp, local features can dominate; where the boundary is ambiguous, the module can lean more heavily on global context to make an informed guess. According to the authors, this dynamic weighting significantly improves the overall feature representation of the pancreas compared with static fusion strategies.

To evaluate the approach, the researchers turned to the NIH public pancreas segmentation dataset, a widely used benchmark assembled from contrast-enhanced abdominal CT volumes with expert-annotated pancreas masks. Performance was measured with the Dice Similarity Coefficient, or DSC, the standard metric in segmentation research, which quantifies the overlap between the predicted organ region and the ground-truth annotation on a scale from zero to one hundred percent. MAF-Net achieved an average DSC of 85.47 percent with a standard deviation of 4.64 percent across the dataset. That average represents a 2.19 percentage point improvement over the baseline models against which it was compared, and the authors state that it outperformed other segmentation models tested in the study. In a field where the pancreas has historically lagged behind easier organs like the liver or kidneys by wide margins, gains of this size are meaningful, and the relatively high standard deviation is itself a reminder of how variable the task remains from patient to patient.

The clinical motivation behind this line of work is sobering. Pancreatic cancer is among the deadliest malignancies, largely because it is frequently diagnosed at an advanced stage, and accurate delineation of the pancreas on imaging is a prerequisite for computer-aided detection of tumors, for planning surgical interventions, and for monitoring disease progression. The references cited in the paper trace a decade of effort on the problem, from early region-growing and threshold-based methods, through cascaded two-stage convolutional networks that first located the organ and then refined its boundary, to recent architectures built on U-Net variants, recurrent neural contextual learning, transformers and attention mechanisms. The new study also situates itself within the current wave of interest in foundation-model approaches, citing recent comparisons between prompt-based Segment Anything Model variants and classical U-Net pipelines for two-dimensional pancreas segmentation, as well as newer networks aimed at joint pancreas and pancreatic tumor segmentation.

What distinguishes MAF-Net within this crowded landscape is its insistence that scale diversity and adaptive fusion be handled explicitly rather than left to the network to discover on its own. Multi-scale feature fusion has appeared in earlier pancreas segmentation work, and attention mechanisms have likewise been used before, but the paper’s contribution lies in packaging parallel multi-scale convolution with spatial attention in a single block and then coupling it to a fusion module that treats local-versus-global weighting as a learned, dynamic decision. The authors trained the network using modern regularization and scheduling techniques, referencing decoupled weight decay and stochastic gradient descent with warm restarts, both standard tools for stabilizing deep network optimization. The research was supported by the Hunan Provincial Natural Science Foundation of China, the Science and Technology Plan Project of Hunan Province, and a Hunan provincial key disciplines program.

For the broader field of medical image analysis, the result is a data point in an ongoing argument about how best to handle small, low-contrast organs in large, cluttered images. One camp favors ever-larger three-dimensional models that reason across entire CT volumes; another favors efficient two-dimensional networks that can be trained on modest datasets and deployed cheaply. MAF-Net belongs to the second tradition, and its reported performance suggests that careful architectural design, rather than sheer model size, still has room to deliver gains. The authors declare no competing interests, and all data supporting the results are included in the article, which should make the claims straightforward for other groups to scrutinize and reproduce on the same public benchmark.

The immediate next question, which the published abstract does not address, is how the network generalizes beyond the NIH dataset, to scans from different scanners, protocols and patient populations, and to pathological abdomens where pancreatitis or tumors distort the anatomy the model was trained to recognize. Previous studies in the cited literature have shown that deep learning methods for pancreas segmentation are sensitive to technical and clinical factors, so validation on diverse external cohorts will be the real test. Still, for a task that has frustrated researchers for over a decade, an 85 percent average Dice score from a purpose-built attention fusion network is a notable milestone, and it adds to the growing evidence that multi-scale context and adaptive attention are becoming indispensable ingredients in the automated reading of abdominal CT scans.

Subject of Research: Deep learning architecture for automated pancreas segmentation in abdominal CT images

Article Title: MAF-Net: multi-scale attention fusion network for pancreas segmentation

Article References: Li, Y., & Deng, H. (2026). MAF-Net: multi-scale attention fusion network for pancreas segmentation. Applied Intelligence, 56(15), Article 472. https://doi.org/10.1007/s10489-026-07514-5

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07514-5

Keywords: pancreas segmentation, deep learning, medical imaging, CT scans, attention mechanism, multi-scale feature fusion, convolutional neural networks, Dice Similarity Coefficient, computer-aided diagnosis, spatial attention, NIH dataset, Applied Intelligence

A team of researchers from Qatar University and the University of Hyogo has now taken a direct swing at that problem. In a study published in Multimedia Tools and Applications, they introduce TomoMamba, a two-stage deep learning framework that, according to the authors, is the first application of selective state-space models to inter-slice feature propagation in DBT. The work arrives at a moment when the Mamba architecture, a relative newcomer to the deep learning scene, has been making waves for its ability to model long sequences with linear computational cost, in contrast to the quadratic cost of transformer attention.

The framework’s design mirrors the clinical workflow of breast screening. Stage 1 acts as a high-throughput triage system: a dual-branch 2.5D convolutional network processes the craniocaudal and mediolateral oblique views of each breast through a shared ResNet-18 backbone, aggregates information across slices with a learnable attention mechanism, and fuses the two view-level representations into a single verdict of normal or suspicious. On the validation partition of the public BCS-DBT dataset, this screening stage achieved a breast-level area under the receiver operating characteristic curve of 0.97, correctly filtering out 94.5 percent of normal breasts while identifying 36 of 41 suspicious ones.

Only breasts flagged as suspicious proceed to Stage 2, where the real architectural novelty lives. A ResNet-18 backbone with a lightweight feature pyramid network extracts features from each of the fifteen representative slices retained per volume. Then, at every spatial position in the feature map, the model extracts a fifteen-element sequence of slice-wise features and feeds it through a Mamba selective state-space model. The crucial distinction from earlier recurrent approaches is that Mamba’s state-transition matrices are input-dependent: the amount each slice contributes to the running hidden state varies with the content of that slice. A lesion that activates strongly on two or three adjacent slices therefore leaves a depth summary that carries information about where along the depth axis the signal appeared, something slice-independent methods discard entirely and full 3D convolutions capture only at substantial memory cost.

The Mamba output is folded back into the per-slice features through a gated residual connection, a sigmoid gate that makes per-channel keep-or-suppress decisions conditioned on the depth context. Introspection across 72 validation volumes showed the gate is channel-selective rather than slice-selective: at lesion positions a mean of 40 of 128 channels fell outside the neutral band, against 23.5 at background positions, a contrast of 1.70 times with a large effect size. Downstream, an anchor-free CenterNet-style detector predicts heatmaps, box sizes and offsets directly from the Mamba-enhanced features, and classification is performed on small feature patches cropped at the detected lesion centers rather than pooled globally. That cropping choice proved essential: a pooling-only baseline oscillated between all-cancer and all-benign predictions without ever converging on the tiny training set of 200 biopsied volumes from just 101 patients.

The evaluation rests on BCS-DBT, hosted on The Cancer Imaging Archive and collected at Duke University Hospital, which remains the only public DBT dataset combining biopsy-confirmed pathology labels with lesion-level bounding-box annotations. The class distribution is extreme, with normal cases outnumbering cancers by more than 100 to 1, which is precisely why the authors split the problem in two. On the validation partition, Stage 2 attained a patient-level AUC of 0.80 with a lesion detection rate of 72.0 percent. On the untouched test partition, however, classification fell to 0.62, an interval that spans chance, while detection degraded only modestly to 62.5 percent. The authors are unusually candid about this: they present lesion localization, not benign-versus-cancer discrimination, as Stage 2’s demonstrated contribution.

Chained together, the two stages reached a breast-level AUC of 0.93 on validation and 0.92 on the held-out test set, forwarding just 14 percent of breasts for expensive diagnostic analysis while losing two of 20 cancers at screening. The authors warn that this cascade figure reflects the screening stage’s normal-versus-suspicious separation rather than genuine cancer discrimination, and that roughly one cancer in ten would be missed if the cascade were used as a standalone filter. They position the system as a reader-support tool to be used alongside, not in place of, radiologist interpretation.

The controlled baseline comparisons are where the architectural argument gets its sharpest test. Under identical data, preprocessing and hyperparameters, TomoMamba was pitted against a bidirectional GRU, a transformer variant and a slice-independent baseline, all matched to within one percent on parameter count. On the validation partition, the cross-slice architectures were not separable by classification AUC, and the authors decline to rank them on that metric. But at matched false-positive rates, TomoMamba localized more lesions than every alternative on validation, and on the test partition it achieved the highest patient-level AUC and the lowest Brier score of the five configurations. Most tellingly, removing the gated cross-slice residual dropped patient-level AUC from 0.798 to 0.548, the only architectural ablation that reached statistical significance at 0.009.

The team also confronted a subtle methodological hazard head-on. Because radiologist annotations guided which slices were retained during preprocessing, the model could in principle have exploited the annotated slice’s position as a hidden marker. The authors tested this directly: detection rates showed no significant difference depending on where the annotated slice sat in the sequence, and under a fully annotation-independent selection procedure, patient-level classification changed little, falling from 0.798 to 0.748 on validation. Since annotations do not exist at deployment, they treat the annotation-independent figures as the deployment-relevant results. Failure analysis added further nuance: the dominant failure mode was classification rather than localization, with most misclassifications clustered within 0.03 of the decision threshold, suggesting a calibration problem rather than a detection one.

The limitations are stated with rare honesty. The study is single-institution, single-scanner, and underpowered for benign-versus-cancer classification with only 101 biopsied training patients; no published DBT method has yet demonstrated cross-site transfer. The much-hyped efficiency of state-space models also did not materialize at this scale: at fifteen slices per volume, Mamba’s sequential scan was actually slower than both the recurrent and attention baselines, though its linear scaling should win out on deeper acquisitions. Still, the significance of the work lies in its disciplined template for evaluating new architectures on small clinical datasets, reporting all three official partitions, quantifying seed variability, and refusing to claim more than the data can support. As larger annotated DBT cohorts emerge, the question of whether selective state-space propagation becomes the backbone of volumetric cancer screening will finally get the data it deserves.

Subject of Research: Deep learning with selective state-space models for cross-slice analysis in digital breast tomosynthesis for breast cancer screening and diagnosis

Article Title: TomoMamba: a two-stage model with state-space cross-slice propagation for breast cancer diagnosis in digital breast tomosynthesis

Article References: Soliman, S., Zafari, Y., Rashed, E. A., & Mabrok, M. (2026). TomoMamba: a two-stage model with state-space cross-slice propagation for breast cancer diagnosis in digital breast tomosynthesis. Multimedia Tools and Applications, 85(10), Article 776. https://doi.org/10.1007/s11042-026-21943-x

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21943-x

Keywords: digital breast tomosynthesis, breast cancer, Mamba, state-space models, deep learning, lesion detection, computer-aided diagnosis, BCS-DBT dataset, medical imaging AI, CenterNet, screening cascade, feature propagation

The clinical motivation is straightforward. Endomicroscopy, which allows in vivo, cellular-level visualization of tissue without a physical biopsy, provides critical pathophysiological information for diagnosing gastrointestinal, pulmonary, and neurological abnormalities. Yet conventional high-resolution endoscopes depend on bulky pixelated detector arrays such as CMOS or CCD sensors, and those arrays cannot be miniaturized down to the hundred-micron scale required to reach highly constrained anatomical regions, including the deep brain, without sacrificing resolution. The result is that some of the most medically valuable imaging sites remain effectively off-limits to high-fidelity, real-time optical imaging.

Optical-fiber-based ghost imaging, also known as single-pixel imaging, has long been considered theoretically ideal for these photon-starved, size-constrained environments. Instead of a camera with millions of pixels, the approach projects a sequence of known light patterns onto a scene and records only the total reflected or transmitted intensity with a single detector. Correlating the known patterns with the measured intensities reconstructs an image computationally. In principle, the optical front end can be as simple as one fiber and one detector, which is exactly what a minimally invasive probe demands.

In practice, however, ghost imaging has been held back by several stubborn technical challenges, chief among them a fundamental speed bottleneck. Conventional implementations require slow, sequential pattern projection, either through spatial light modulators that display one pattern at a time or through wavelength sweeps that step through illumination colors sequentially. That serial approach caps the achievable frame rate and introduces severe motion artifacts whenever the target moves, which is precisely the situation in living tissue. Classical reconstruction algorithms add a second problem: substantial computational latency that makes real-time, video-rate image reconstruction impractical for clinical use.

The new system resolves both bottlenecks through a hardware-software co-design. On the hardware side, the researchers leveraged highly stabilized dual optical frequency combs, devices that generate spectra composed of many precisely spaced, phase-coherent lines. Each individual comb line is mapped to a mutually uncorrelated speckle pattern, meaning that many distinct illumination patterns are effectively generated and detected in parallel rather than one after another. Wavelength-division multiplexing, the same technology that carries many data channels through a single optical communications fiber, combines with dual-comb interferometry and ghost imaging to create a minimally invasive optical front end.

The detection step is equally elegant in its compression. After the comb light interacts with the target, the system optically measures the bucket sum of each comb-line power, essentially a hyperspectral multiply-accumulate, or MAC, operation performed in parallel across the spectral channels. In a single photodetection step, the hypercube of data containing two-dimensional spatial information is physically compressed into a one-dimensional analog temporal electrical voltage signal. All of the spatial detail survives only implicitly, encoded in how each speckle pattern weighted the light that reached the detector.

Recovering an image from such a highly compressed snapshot is fundamentally an inverse optimization problem, and this is where the software half of the co-design takes over. The team developed an application-specific deep-learning Transformer model that learns the complex correlations between the known speckle patterns and the corresponding bucket intensities of the target-encoded light. Once trained, the model reconstructs high-fidelity target images from the compressed measurements, outperforming classical reconstruction algorithms in both raw imaging accuracy and processing speed. The combination of parallel optical pattern generation and fast learned reconstruction is what pushes the system past video rates for dynamic scenes.

Dr. Myoung-Gyun Suh, Senior Scientist and Group Head at NTT Research’s Physics and Informatics Laboratories and lead author of the study, emphasized the significance of the combined approach. By merging the inherent parallelism and precision of optical frequency combs with the reconstruction capabilities of a Transformer-based deep learning model, he noted, the team eliminated the sequential projection bottleneck and significantly improved imaging speed and fidelity. He added that the hardware-software co-design dramatically simplifies the optical front end, making high-fidelity dynamic ghost imaging highly viable for size-constrained scenarios such as single-use endomicroscopic probes.

The implications extend well beyond the laboratory bench. Because the entire optical front end consists of a single-core fiber and a single-pixel detector, the architecture is naturally compatible with the narrow, flexible probes that endomicroscopy requires, and one example application highlighted by the researchers is neurosurgical endomicroscopy, where a minimally invasive optical probe must reach deep, confined anatomy. The same properties also suit portable sensing in other constrained environments where cameras and scanning optics cannot fit. By demonstrating the concept experimentally on dynamic targets, the team has moved ghost imaging a significant step toward practical deployment in real-world biomedical applications, freeing the technique from the optical table and bringing it closer to clinical neurosurgical and diagnostic use.

Subject of Research: Dual-comb ghost imaging with deep learning reconstruction for minimally invasive fiber-based endomicroscopy

Article Title: Breakthrough in optical-fiber-based minimally invasive imaging using optical frequency combs and deep learning

Article References: Breakthrough in optical-fiber-based minimally invasive imaging using optical frequency combs and deep learning. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: ghost imaging, optical frequency combs, deep learning, Transformer model, endomicroscopy, single-pixel imaging, dual-comb interferometry, wavelength-division multiplexing, hyperspectral imaging, computational imaging, biomedical optics, minimally invasive imaging