To understand why this matters, it helps to consider what segmentation networks are asked to do. A U-Net, the workhorse architecture of medical image analysis since 2015, consists of an encoder that progressively compresses an image into abstract feature maps and a decoder that expands those maps back to full resolution while predicting, pixel by pixel, whether each location belongs to tumor or healthy tissue. The classic weakness of this design lies in the handshake between the two halves. Encoder features carry rich global semantics but coarse spatial detail, while decoder features preserve fine boundaries but can lose sight of the larger context. Conventional U-Net variants simply concatenate or add these feature sets, a blunt operation that leaves the network to figure out on its own which information from each side deserves attention.

The SECA module addresses this gap directly. Positioned in the decoder, it explicitly models cross-attention between encoder and decoder features, learning to weigh which global semantic cues should guide the alignment of fine-grained spatial details at each scale. In practical terms, the network can decide, for example, that a faint density pattern detected by the encoder in a mammogram is highly relevant to sharpening an ambiguous boundary in the decoder’s current prediction, and it can propagate that relevance systematically rather than leaving it to chance. The authors emphasize that this attention-driven decoding, rather than the mere use of a Swin Transformer encoder, is the primary contribution of their framework. The Swin encoder itself contributes the ability to capture long-range dependencies and global contextual information through its shifted-window attention mechanism, which computes relationships between image patches efficiently even at high resolutions typical of clinical images.

A second challenge the researchers tackled is the sheer variability of breast tumors. Lesions can be tiny specks or large irregular masses, with shapes ranging from well-circumscribed ovals to spiculated forms that infiltrate surrounding tissue. A decoder operating at a single scale tends to favor tumors whose size matches its receptive field. To counter this, the team integrated a transformer-guided multi-scale decoder based on Inception modules, an architectural pattern that processes features through parallel branches of different sizes within the same layer. This allows the network to simultaneously examine a region at multiple zoom levels, so that a small mass and a sprawling lesion can both be segmented accurately within the same forward pass, and the transformer guidance ensures that the multi-scale branches remain informed by the global context captured upstream.

Training such a network also demanded a careful loss function, the mathematical objective that tells the model how wrong its predictions are. Segmentation of breast tumors suffers from class imbalance, since tumor pixels typically occupy only a small fraction of each image, meaning a naive model could score well by predicting that nothing is cancerous. The proposed framework adopts a hybrid loss combining three components. The Dice loss measures overlap between predicted and true tumor regions, directly optimizing the metric clinicians care about. The Focal loss focuses the model’s learning effort on hard samples, the ambiguous pixels that a standard loss would let it gloss over. The Boundary loss concentrates on the tumor perimeter, where delineation precision matters most for measuring lesion size and planning surgery or biopsy. Together, these three terms push the network toward accurate, confident, and cleanly outlined segmentations.

The evaluation strategy is one of the study’s strongest points. Rather than testing on a single dataset, the authors benchmarked the model on four public datasets spanning the full range of breast imaging: CBIS-DDSM for mammography, BreastDM for dynamic contrast-enhanced MRI, BCSS for histopathology whole-slide images, and the Breast Ultrasound Images dataset. Each modality presents distinct obstacles. Mammograms are low-contrast and cluttered with dense fibroglandular tissue. MRI volumes capture how tumors enhance with contrast agent over time. Histopathology slides reveal cellular textures at extreme resolution. Ultrasound images are noisy with speckle artifacts and shadowing. An architecture that performs well across all four is demonstrating genuine generalizability rather than overfitting to the quirks of one imaging technique.

The results showed consistent performance gains over established baselines, including the original U-Net, the nested U-Net++ variant, and existing Swin U-Net architectures. Improvements reached up to 2.1 percentage points in Dice score, the standard overlap measure for segmentation quality, and up to 1.7 percentage points in Intersection over Union, a related metric that penalizes both missed tumor regions and false alarms. In a field where incremental gains of a fraction of a point are common, improvements of this magnitude across four heterogeneous datasets are notable. Ablation studies, in which individual components are removed one at a time, confirmed that the SECA module, the multi-scale Inception-based decoder, and the hybrid loss each contribute measurably to the overall performance, validating that the gains stem from the design rather than from added parameters alone.

Qualitative analyses reinforced the quantitative findings, with the model producing segmentations that adhere more closely to tumor boundaries and generate fewer spurious regions than competing methods, particularly in images where tumors blend into surrounding tissue. The authors have committed to releasing the full implementation, including the SECA module, the hybrid loss function, training scripts, and evaluation protocols, through a public GitHub repository, a transparency measure that should allow other research groups to reproduce the results and adapt the architecture to new clinical contexts.

The broader significance of this work lies in its potential to standardize and accelerate a task that currently depends heavily on radiologist and pathologist expertise. Manual tumor delineation is time-consuming, and inter-observer variability between experts is well documented, meaning that two specialists examining the same scan may outline a tumor differently. A robust automated segmentation tool that performs reliably across modalities could serve as a consistent second reader, flagging lesions, pre-measuring tumor extent, and reducing the cognitive load on clinicians during screening and diagnosis. Because the architecture is modality-agnostic in its design principles, the SECA cross-attention strategy and multi-scale decoding could plausibly transfer to other organs and other segmentation problems where encoder-decoder alignment and scale variability are limiting factors.

Challenges remain before such tools reach routine clinical use. Prospective validation on real-world hospital data, regulatory review, and integration into picture archiving and communication systems all lie ahead, and the study itself, published in September 2026 in Multimedia Tools and Applications, was conducted without external funding and reports no conflicts of interest. Nevertheless, by attacking the specific technical weaknesses of encoder-decoder communication and multi-scale tumor representation, and by proving the approach across mammography, MRI, histopathology, and ultrasound in a single framework, the researchers have moved the field a meaningful step closer to AI-assisted breast cancer diagnosis that is accurate wherever the tumor appears and whatever the camera that captured it.

Subject of Research: Deep learning-based breast tumor segmentation across multiple medical imaging modalities

Article Title: A multi-scale transformer U-Net with SECA attention for robust breast tumor segmentation across diverse imaging modalities

Article References: Ahmadi Lashaki, R., Bayrami, F., Rokhva, S., & Aghdaei, E. (2026). A multi-scale transformer U-Net with SECA attention for robust breast tumor segmentation across diverse imaging modalities. Multimedia Tools and Applications, 85(9), Article 736. https://doi.org/10.1007/s11042-026-21810-9

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21810-9

Keywords: breast cancer, tumor segmentation, Swin Transformer, U-Net, cross attention, medical imaging, deep learning, mammography, MRI, histopathology, ultrasound, hybrid loss function

The research, published in Veterinary Oncology, focused on canine acanthomatous ameloblastoma, or CAA, the most common odontogenic tumor in dogs, accounting for roughly ten percent of oral neoplasms in a species where oral tumors overall represent six to seven percent of all cancers. The gold standard treatment is surgical excision with margins of five to ten millimeters, which makes accurate tumor delineation on preoperative imaging critically important. If the tumor boundary drawn on a CT scan is wrong, the surgeon’s plan may be wrong too, and incomplete excision carries well-documented consequences: higher recurrence rates, reduced survival, and poorer overall prognosis.

To test the reliability of manual tumor outlining, the team selected ten confirmed CAA cases from the Cornell University Hospital for Animals CT database covering 2014 to 2021. Every case had been diagnosed by histopathology, and each scan included both non-contrast and contrast-enhanced images with bone and soft tissue reconstructions. The cohort comprised four Labrador retrievers and six other dogs of assorted breeds, with a mean age of 7.4 years. Nine tumors sat in the mandible, six rostrally and three caudally, while one arose from the left maxillary canine. Scans were acquired on a Toshiba Aquilion 16-slice scanner at 120 kilovolt peak, with pixel dimensions averaging 0.40 millimeters and slice thicknesses ranging from 0.5 to 2 millimeters.

Three board-certified specialists, two veterinary dentists and oral surgeons and one veterinary medical oncologist, independently contoured each of the ten tumors twice, in two randomized sittings separated by at least one week to reduce recall bias. Working in the open-source platform 3DSlicer, they could toggle between contrast and non-contrast datasets and use paintbrush or drawing tools on each transaxial slice. They received only crude tumor locations and a dental map, with no radiology reports or medical records, deliberately mimicking the conditions under which such contouring is often performed in practice without radiologist input. Each case took between 30 and 60 minutes, and no participant faced time restrictions.

The investigators then compared all the resulting contours using three complementary metrics. The Dice Similarity Index quantifies spatial overlap between two segmentations, ranging from zero for no overlap to one for perfect overlap. The Hausdorff Distance captures the worst-case error, measuring the maximum distance between the boundaries of two contours. Centroid positions, the arithmetic mean coordinates of all voxels in a segmented volume, reveal whether observers placed the tumor’s geometric center in the same spot. Volumes were also compared, with the median tumor volume across all trials measuring 2.46 cubic centimeters and a range from 1.01 to 9.46 cubic centimeters.

The headline finding was a striking split between global and local agreement. Correlation coefficients for centroid coordinates and tumor volumes showed excellent agreement both within and between observers, suggesting that everyone was capturing roughly the same mass in roughly the same place. Yet the boundary-level metrics told a more nuanced story. Intraobserver Dice values averaged 0.78 with a standard deviation of 0.11, indicating only moderate agreement even with oneself across repeated trials. Interobserver Dice values dropped to a mean of 0.69 with a standard deviation of 0.14, which the study’s classification scheme labels as poor agreement between different participants.

Hausdorff Distances reinforced that pattern. Within a single observer’s two trials, the mean worst-case boundary discrepancy was 0.80 millimeters, but between different observers it climbed to 1.21 millimeters, with a standard deviation of 0.66. Effect size analysis using Cohen’s d showed small, medium, and large effects for the three pairwise observer comparisons, and interobserver volume estimates were statistically different at P less than 0.001, with effect sizes ranging from 0.31 to 0.72. Analysis of variance also flagged significant differences in the centroid Z coordinate between one pair of observers and in volumes between another. Notably, the authors estimated that observer differences accounted for 15.6 percent of volume variability and 17.7 percent of Hausdorff Distance variability.

Why does this matter beyond veterinary clinics? Because manual segmentations are the training data for the radiomics and AI models now being built to diagnose, classify, and prognose tumors automatically. In human medicine, variations in gross tumor volume delineation can reach as much as twenty percent, and radiomics-based models have shown good diagnostic performance in head and neck cancers, canine small intestinal tumors, and renal tumors. If the ground truth used to train these algorithms is itself inconsistent, the models inherit that inconsistency. The Cornell team frames their work as a technical validation step, one of the internal, external, and technical validation processes required before AI models can be trusted in the clinic, and their reliability data can now serve as a benchmark for comparing segmentation methods and developing autosegmentation tools.

The authors are candid about their study’s limitations. Ten cases and three observers is a small sample, CT acquisition parameters varied across cases, slice thickness ranged up to 2 millimeters along the Z axis, and the cohort overrepresented rostral mandibular tumors and Labrador retrievers, factors that could matter when building radiomics models even if they likely had little effect on observer agreement itself. Time per case was self-reported rather than measured, so learning effects cannot be excluded, and the design deliberately departed from standard clinical workflow by withholding radiology and pathology reports. The team also notes that whether these minor contour inconsistencies would actually change surgical margins or outcomes cannot be determined without histopathological assessment of complete excisions, a question they hope larger studies will address.

Still, the takeaway is clear and broadly relevant: even experienced specialists tracing the same tumor on the same scans produce boundaries that overlap only moderately, even though the tumor’s center and size come out remarkably consistently. As radiomics and AI move from promising research tools toward clinical integration in both human and veterinary oncology, the study argues that standardized contouring protocols are no longer optional. Reliable segmentation, the authors conclude, is foundational, and quantifying exactly how much human hands disagree is the first step toward building machines that do it better.

Subject of Research: Observer variability in manual CT tumor delineation for canine acanthomatous ameloblastoma and its implications for radiomics validation

Article Title: Assessment of inter and intraobserver variability in manually delineated oral canine acanthomatous ameloblastoma from computed tomography scans

Article References: Drozd, M. E., Fiani, N., Sylvester, S., Khalil, M. K., Peralta, S., & Basran, P. S. (2025). Assessment of inter and intraobserver variability in manually delineated oral canine acanthomatous ameloblastoma from computed tomography scans. Veterinary Oncology, 2(1), Article 37. https://doi.org/10.1186/s44356-025-00052-1

Image Credits: AI Generated

DOI: 10.1186/s44356-025-00052-1

Keywords: radiomics, canine acanthomatous ameloblastoma, computed tomography, tumor segmentation, observer variability, Dice Similarity Index, Hausdorff distance, veterinary oncology, 3DSlicer, artificial intelligence, surgical margins, gross tumor volume

The case for automation begins with a measurement problem. Lymphoma is a highly heterogeneous blood cancer that can seed lesions virtually anywhere in the body, and the standard treatment of visual reading of whole-body PET scans is slow and notoriously variable between observers. One of the most powerful prognostic markers in the field, total metabolic tumor volume, or TMTV, requires clinicians to trace the boundaries of every FDG-avid lesion across an entire scan. Manual delineation is so labor-intensive and subjective that it has historically blocked the routine clinical adoption of this biomarker, despite extensive validation across diffuse large B-cell lymphoma, Hodgkin lymphoma, peripheral T-cell lymphoma, and even patients heading into CAR-T cell therapy. Traditional computer-assisted methods such as thresholding, region growing, and edge detection helped but never fully solved the reproducibility problem.

Deep learning has changed that equation. Convolutional neural networks, particularly architectures built on the three-dimensional U-Net template, can now learn directly from images how to outline tumor tissue without hand-crafted rules. The review highlights a wave of refinements tailored to lymphoma’s specific challenges. Dense U-Net variants with Tversky loss functions sharpen performance on very small lesions, which are commonly missed by generic models. Prior-shift regularization networks tackle the extreme class imbalance of whole-body scans, where tumor occupies a tiny fraction of the pixels. A longitudinally aware segmentation network tracks lesions across serial scans in pediatric Hodgkin lymphoma, improving interim treatment monitoring without sacrificing baseline accuracy. Boundary-optimized and cruciform-guided designs have pushed Dice similarity coefficients as high as 90.7 percent in some studies, while hybrid PET/CT feature fusion modules have achieved detection accuracies exceeding 99 percent in body-region-level detection tasks.

Crucially, the field is proving that these tools generalize. TMTV-Net, a cascaded 3D U-Net system, achieved comparable performance on internal test data and on independent multi-site external data, with average Dice scores around 0.66 to 0.68, a critical benchmark for regulatory and clinical credibility. Studies in diffuse large B-cell lymphoma demonstrated that fully automatic TMTV calculation from U-Net segmentations serves as an independent survival predictor, and a fully automated pipeline combining imaging metrics with clinical variables outperformed the International Prognostic Index, the field’s long-standing standard, for progression-free and overall survival prediction. In a striking analysis of 1,268 patients from the phase III GOYA trial, the automated tool even flagged a high-risk subgroup for central nervous system relapse, a rare but devastating metastatic pattern that baseline imaging had not previously been leveraged to anticipate.

Because supervised deep learning normally demands enormous annotated datasets, researchers have also turned to cleverer training strategies. Semi-supervised approaches that blend classical fuzzy clustering loss functions with U-Net training matched or beat fully supervised baselines using limited labels. Transfer learning frameworks trained on limited annotations generalized across six different cancer types, including lymphoma, and generative adversarial networks that synthesize metabolically enhanced images boosted the performance of independent segmentation models. Even the scans themselves are being compressed: a self-attention U2Net framework cut PET acquisition time by 50 percent while preserving diagnostic quality, and a deep neural network classified FDG PET/CT scans with and without hypermetabolic tumor sites with an area under the curve of 0.95 in a dual-center retrospective analysis.

The second frontier is differential diagnosis, where machine learning models are moving PET/CT from a staging tool toward a non-invasive diagnostic instrument. One of the most clinically consequential applications is predicting bone marrow involvement, which currently requires invasive biopsy and is vulnerable to sampling error. An interpretable multicenter model combining clinical data, PET parameters, radiomic features, and deep learning features achieved an area under the curve of 0.886, with explainability analysis identifying PET radiomic signatures, platelet count, and B symptoms as the dominant predictors. Such tools could spare many patients an unnecessary bone marrow procedure while catching cases that biopsy misses.

Subtype classification follows a similar logic. Because aggressive and indolent lymphomas demand fundamentally different treatment strategies, and because a single needle biopsy may not capture spatially heterogeneous disease, imaging-based classifiers offer a complementary window. A pseudo spatial-temporal radiomics approach exploits intratumoral metabolic heterogeneity by simulating recurrent network behavior across multi-threshold tumor volumes, outperforming conventional radiomics in tumor classification. An ensemble model built on super-resolution imaging predicted lymphoma type with 94.8 percent accuracy, while a hybrid few-shot multiple-instance learning framework distinguished aggressive diffuse large B-cell lymphoma from indolent follicular lymphoma with an area under the curve of 0.795 despite limited annotated training data. Beyond lymphoma itself, deep convolutional networks separated sarcoidosis from lymphoma on maximum-intensity-projection PET images with an area under the curve of 0.963, radiomics-guided machine learning matched or exceeded physician performance in the same discrimination task across more than 400 patients, and computer-aided models distinguished cervical lymph node metastasis from lymphoma involvement with an area under the curve of 0.901. Machine learning models have even begun separating mass-forming pancreatic lymphoma from pancreatic carcinoma, one of the trickiest diagnostic mimics in oncology imaging.

Prognostic stratification, however, is where the authors see the deepest transformation. Conventional tools, including the International Prognostic Index, the NCCN-IPI, and the five-point Deauville score, capture only crude dimensions of tumor biology. AI models now extract high-throughput textural features, spatial dissemination patterns, and deep image signatures that quantify heterogeneity the human eye cannot grade. In diffuse large B-cell lymphoma, multi-timepoint radiomics models combining baseline and end-of-treatment scans beat traditional indices at predicting two-year progression, and deep learning scores extracted from standard network architectures demonstrated robust stability in multicenter external validation. The field is also fusing modalities: one prognostic index merged deep learning PET biomarkers with genetic subtypes to offer regimen-specific risk discrimination, and automated machine learning pipelines built multimodal PET/CT deep feature signatures for survival prediction in elderly patients. Graph neural networks, which convert PET scans into attributed lesion graphs fused with clinical data through cross-attention mechanisms, are emerging as a leading answer to the black-box problem, identifying which specific lesions drive high-risk predictions while maintaining accuracy.

The review is candid that progress is uneven across subtypes and settings. In Hodgkin lymphoma, machine learning models predicted two-year event-free survival with an area under the curve of 0.81, and a low-cost algorithm combining routine blood tests offers interim response assessment for children in regions without PET scanner access. In extranodal NK/T-cell lymphoma, a deep learning system integrating segmentation, image fusion, and prognosis yielded reliable survival predictions in multicenter cohorts, while weakly supervised approaches compensated for missing follow-up data. Yet a sobering counterexample from a prospective trial with PET-guided consolidation strategies found that radiomic features added no significant predictive gain, a reminder that the value of any AI model depends on the clinical context and existing interventions around it.

The path to routine clinical use now runs through validation, standardization, and trust. Only four of the 53 reviewed studies were prospective, and just 20 included independent external validation, leaving most models exposed to selection bias and local protocol overfitting. The authors call for standardized imaging protocols, federated learning to navigate privacy regulations, seamless integration with hospital information and picture archiving systems, rigorous regulatory pathways appropriate for Class III medical devices, and explainable AI architectures that map imaging features onto biological mechanisms. If those conditions are met, the trajectory is clear: PET/CT-based AI is evolving from a manual feature-extraction exercise into automated, interpretable, multimodal decision support that, combined with genomics and longitudinal clinical data, could finally deliver the personalized, precision management that lymphoma’s extraordinary heterogeneity has always demanded.

Subject of Research: Artificial intelligence-based analysis of FDG PET/CT imaging for automated lesion segmentation, differential diagnosis, and prognostic stratification in lymphoma

Article Title: Artificial intelligence-based PET/CT analysis in lymphoma: segmentation, differential diagnosis, and prognostic stratification

Article References: Liu, Q., Sun, R., Chen, X., & Liu, Y. (2026). Artificial intelligence-based PET/CT analysis in lymphoma: segmentation, differential diagnosis, and prognostic stratification. Holistic Integrative Oncology, 5(1), Article 73. https://doi.org/10.1007/s44178-026-00294-5

Image Credits: AI Generated

DOI: 10.1007/s44178-026-00294-5

Keywords: lymphoma, PET/CT, artificial intelligence, deep learning, radiomics, tumor segmentation, total metabolic tumor volume, prognostic stratification, differential diagnosis, convolutional neural networks, graph neural networks, precision medicine