The challenge the researchers set out to solve is deceptively simple to describe but notoriously difficult in practice. CTA scanners produce three-dimensional volumetric images of the head in which contrast-enhanced blood vessels glow brightly against surrounding tissue. An aneurysm appears as an abnormal outpouching of a vessel wall, but these lesions are often tiny, sometimes only a few millimeters across, and their boundaries blend gradually into the parent artery. A single CTA scan contains hundreds of slices, and radiologists must visually comb through the entire vascular tree, where an aneurysm may hide behind bone, overlap with adjacent vessels, or mimic normal anatomical variants. Missed aneurysms can be fatal, while false alarms trigger unnecessary invasive investigations. Automated segmentation promises to relieve this burden, yet most existing algorithms struggle precisely where it matters most: at the ambiguous boundary between a small aneurysm and the healthy vessel from which it arises.

HSF-Net builds on a modern backbone architecture known as Swin UNETR, which combines the strengths of two competing paradigms in medical image analysis. Convolutional neural networks, the workhorses of computer vision, excel at capturing local textures and fine spatial detail, while transformer networks, originally developed for language processing, capture long-range dependencies by letting every part of an image attend to every other part. The Swin Transformer variant achieves this efficiently by computing attention within shifted local windows, making it computationally feasible for the enormous three-dimensional volumes that CTA produces. Swin UNETR had already proven effective for brain tumor segmentation in magnetic resonance images, but the Chinese team recognized that its decoder, the half of the network responsible for reconstructing fine-grained segmentation masks from compressed high-level features, relied mainly on simple concatenation of features. For aneurysms, where the critical information lives at multiple scales simultaneously, that design left performance on the table.

The core innovation of HSF-Net lies in what the authors call a Hybrid-Scale Fusion module, inserted into the decoder alongside a Convolutional Block Attention Module, or CBAM. The guiding philosophy is summarized in the paper as fuse first, refine later. Conventional decoders take features from different levels of the network, which encode coarse context and fine detail respectively, and simply stack them together before processing. HSF-Net instead performs an input-dependent fusion in which the network learns, for each individual case, how to weight and combine multi-scale features before any refinement occurs. This matters because aneurysms vary enormously in size, shape, and location; a small saccular bulge at the junction of two arteries demands different feature weighting than a large, irregularly shaped lesion. By letting the fusion process adapt to the input, the network avoids the one-size-fits-all compromise that fixed fusion schemes impose.

Once the hybrid-scale fusion has produced a richly combined feature representation, CBAM takes over the refinement stage. CBAM is a lightweight attention mechanism that operates along two complementary dimensions. Its channel attention component learns which feature channels, essentially which learned pattern detectors, are most relevant to the task at hand, amplifying informative ones and suppressing noise. Its spatial attention component then highlights where in the image those relevant features matter, effectively teaching the network to focus its computational gaze on vessel boundaries and lesion candidates. The sequential application of channel and spatial attention after fusion means that the network first decides what information matters and then decides where it matters, a two-step filtering process that proved especially valuable for delineating the faint, ambiguous edges that make aneurysm segmentation so treacherous.

The empirical results support the design. On the internal independent test set, drawn from clinical CTA data collected at Yijishan Hospital of Wannan Medical College, HSF-Net reached a mean patient-level Dice score of 83.4 percent, outperforming the Swin UNETR baseline and achieving the top score among all evaluated methods under an identical evaluation protocol. The Dice score, a standard metric in segmentation research, measures the overlap between the algorithm’s predicted aneurysm region and the ground truth annotated by experts, with 100 percent representing perfect agreement. Ablation studies, in which the researchers systematically removed individual components, confirmed that both the Hybrid-Scale Fusion module and the CBAM attention mechanism contributed complementary gains, demonstrating that the improvements were not attributable to a single trick but to the interplay of the two designs.

Perhaps the most consequential finding for clinical translation is the external validation. The team evaluated the model on an independent cohort from Wuhu Second People’s Hospital, a separate institution with its own scanners, imaging protocols, and patient population. Deep learning models in medicine routinely suffer from domain shift, a phenomenon in which performance degrades when the data distribution changes between the training hospital and the deployment site. The external evaluation provided preliminary evidence of cross-center applicability, suggesting that HSF-Net’s learned features capture genuine anatomical and pathological structure rather than quirks of a single scanner or annotation style. The authors caution that this remains preliminary evidence, and broader multi-center trials will be needed before such a system could support radiologists in daily practice, but the direction is encouraging.

The clinical stakes of this work are considerable. Unruptured intracranial aneurysms are common, with systematic reviews estimating prevalence in the range of a few percent of the adult population, and their management hinges on accurate assessment of size, shape, and location, which in turn depends on precise segmentation of the lesion in three-dimensional imaging. A reliable automated tool could serve as a second reader, flagging subtle lesions that a fatigued radiologist might overlook, standardizing measurements used in treatment planning for surgical clipping or endovascular coiling, and accelerating workflows in busy stroke centers. The study also used ethically approved clinical data from both cohorts, with informed consent requirements waived, and the underlying data are available from the corresponding author upon reasonable request, reflecting growing norms around reproducibility in medical AI research.

Architecturally, HSF-Net also represents a broader trend in the field: the move away from purely convolutional designs such as the celebrated U-Net family toward hybrid transformer-convolutional systems, and away from static feature combination toward adaptive, attention-guided processing. The lineage is visible in the paper’s references, which trace the evolution from U-Net and V-Net through nnU-Net and UNETR to the Swin Transformer and attention modules like squeeze-and-excitation networks and CBAM itself. What distinguishes HSF-Net is not any single component but the deliberate sequencing of operations, fusing multi-scale evidence before applying attention-based refinement, which the ablation experiments show is more than the sum of its parts. As automated aneurysm analysis matures, designs of this kind could become standard components of the radiology reading room, quietly scanning every CTA study for the small, dangerous bulges that human eyes, however trained, can miss.

Subject of Research: Deep learning segmentation of intracranial aneurysms in CTA images

Article Title: HSF-Net: a hybrid-scale fusion network for intracranial aneurysm segmentation in CTA images

Article References: Yang, S., Wang, P., Ye, M., Liu, L., Li, Y., & Zhou, Y. (2026). HSF-Net: a hybrid-scale fusion network for intracranial aneurysm segmentation in CTA images. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03684-z

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03684-z

Keywords: intracranial aneurysm, CTA, deep learning, image segmentation, Swin UNETR, attention mechanism, multi-scale feature fusion, medical imaging, subarachnoid hemorrhage, neuroradiology, artificial intelligence, HSF-Net

The clinical standard for describing intervertebral disc degeneration is the Pfirrmann grading system, a five-level scale originally devised for lumbar discs and later adapted for the cervical spine. A reader looks at the disc on a T2-weighted image and assigns a grade from I to V based on the loss of the bright nuclear signal, the blurring of the boundary between nucleus and annulus, and the collapse of disc height. The system is widely used and reasonably reproducible when performed by experienced readers, but it is ordinal rather than continuous, which means it cannot capture subtle progression, and it compresses a rich spectrum of tissue change into five coarse bins. Two discs with visibly different appearances can receive the same grade, and the same disc can receive different grades from different readers at different times.

The Suzhou team’s answer was to build a pipeline that begins with a single human decision: an experienced radiologist selects the midsagittal slice, the midline image through the spine where the cervical discs are best profiled. From that point onward, everything is automated. A two-dimensional Swin-UNETR, a transformer-based deep learning architecture, segments four structures on the chosen slice: the vertebral bodies, the intervertebral discs, the cerebrospinal fluid, and the spinal cord. Once these masks exist, the software extracts quantitative indices from each disc, including the relative signal intensity of the disc compared with a reference, a disc height index derived from a Euclidean distance transform, and a height-to-diameter ratio that captures disc shape independently of absolute scale.

The choice of a Swin-UNETR reflects broader trends in medical image analysis. Transformer architectures, which process images through attention mechanisms rather than purely local convolutional filters, have proven adept at capturing long-range context, which matters when the task is to distinguish a disc from the adjacent bone marrow and cerebrospinal fluid on a single grayscale slice. The model was trained and tested entirely within a single institution, and its performance was evaluated on an internal test cohort of 54 patients. The Dice coefficients, a standard overlap metric where 1.0 indicates perfect agreement between automated and reference masks, came in at 0.912 for vertebral bodies, 0.900 for intervertebral discs, and 0.821 for cerebrospinal fluid. For the discs, the structure that matters most for the downstream measurements, that level of agreement is respectable, though the lower score for cerebrospinal fluid hints at the difficulty of segmenting thin, low-contrast fluid spaces.

Reliability of the measurements themselves was assessed in a spectrum-enriched nested subset of 34 patients comprising 170 cervical discs, where automated outputs were compared against reference measurements. When the same distance-transform definition was applied to both automated and reference masks on the physical scale, the disc height index showed a correlation of r = 0.894 and an intraclass correlation coefficient of 0.863, indicating strong agreement. The picture was more complicated for measurements taken from the stored historical image series: the intraclass correlation coefficients were 0.770 for relative signal intensity, 0.635 for the disc height index, and just 0.381 for the height-to-diameter ratio. That last figure is a sobering reminder that signal intensity and geometry measured from routine clinical archives carry variability that can erode reproducibility, a problem that any clinical deployment would need to confront through standardized acquisition protocols.

To establish that the quantitative indices actually track the degenerative process, the researchers turned to a separate spectrum-enriched subset of 30 patients in which two readers independently annotated the discs and assigned Pfirrmann grades. Inter-reader agreement was strong, with a Dice of 0.873 for disc segmentation and a linearly weighted kappa of 0.901 for the Pfirrmann grades, a level of concordance that lends credibility to the grades used as the reference construct. The team then examined how the three quantitative indices related to those grades using patient-clustered statistical analyses adjusted for age, sex, and cervical level, an important step because discs at different cervical levels and in patients of different ages have systematically different baseline appearances.

The associations were consistent in direction. All three indices declined as Pfirrmann grade increased: relative signal intensity showed a Spearman correlation of −0.445 with grade, the stored disc height index −0.384, and the stored height-to-diameter ratio −0.635. After adjustment for age, sex, and level, the overall effect of grade remained statistically significant for all three indices, with P values of 0.00164 for relative signal intensity, 0.0185 for the disc height index, and less than 0.001 for the height-to-diameter ratio. But when the researchers applied the Holm correction for multiple comparisons to the pairwise differences between adjacent grades, only the height-to-diameter ratio retained significant discrimination, and only between grades III and IV and between grades IV and V. In other words, the shape-based index was the most robust at telling neighboring grades apart, while the others captured the overall trend without cleanly separating each step.

The team also tested whether the quantitative features could classify discs into Pfirrmann grades automatically, an exploratory exercise with instructive results. Among the classifiers evaluated, a Random Forest performed best numerically, achieving an accuracy of 0.488 on the five-class task and 0.665 on a simplified three-class version. A Decision Tree model was retained for its interpretability and reached 0.435 and 0.635 respectively. These numbers make clear that five-way grading from a handful of image-derived features remains a hard problem; a continuous degeneration score, rather than a categorical grade, may ultimately be the more natural output for quantitative imaging. The authors are careful to frame the entire classification analysis as exploratory and internally evaluated only.

Indeed, the study’s most admirable quality may be its restraint. The authors state explicitly that the observed associations represent correspondence with an imaging construct rather than independent biological or clinical validation, that all analyses were internal with no external scanner or institutional cohort evaluated, and that further multicenter and multi-scanner validation is required before the framework could be considered for clinical use. This matters because the gap between a promising internal evaluation and a deployable clinical tool is wide. Scanners from different manufacturers, coils, field strengths, and sequence parameters all alter signal intensities and apparent geometry, and a model trained on one center’s archive may not transfer gracefully to another’s. The modest reproducibility of the height-to-diameter ratio on stored historical series is a concrete example of exactly the kind of vulnerability that external validation would expose.

Still, the direction of travel is clear and, for patients with neck pain, potentially consequential. Degenerative cervical myelopathy, the most common cause of spinal cord dysfunction in adults worldwide, often develops silently as discs degenerate, narrow, and impinge on the cord, and earlier, more sensitive detection of disc deterioration could shift diagnosis toward a stage where intervention is simpler and more effective. A framework that turns each disc into a set of continuous, trackable numbers could allow clinicians to follow progression over time with far more sensitivity than a five-point grade allows, and could give researchers the measurement precision needed to test whether new therapies actually slow degeneration. The Suzhou study does not prove that vision, and it is honest enough to say so. What it does provide is a carefully documented demonstration that AI-assisted segmentation can extract physically meaningful, grade-correlated measurements from the images that hospitals already acquire every day, which is precisely the foundation on which the necessary multicenter trials can now be built.

Subject of Research: AI-based quantitative MRI assessment of cervical intervertebral disc degeneration

Article Title: Segmentation-based quantitative characterization of cervical disc degeneration on midsagittal T2-weighted MRI: a retrospective internal evaluation

Article References: Li, X., Gong, C., Tang, Z., Li, Z., Xu, K., Chang, W., Ma, Z., Liu, J., Dai, Y., Li, Y., & Yu, P. (2026). Segmentation-based quantitative characterization of cervical disc degeneration on midsagittal T2-weighted MRI: a retrospective internal evaluation. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02871-6

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02871-6

Keywords: cervical spine, disc degeneration, MRI, deep learning, Swin-UNETR, Pfirrmann grading, image segmentation, quantitative imaging, radiology, spine, machine learning, medical imaging

The study, published in the International Journal of Computer Assisted Radiology and Surgery by researchers from Eindhoven University of Technology and Catharina Hospital, tackles a problem that has long frustrated radiologists. When they examine CT scans for peritoneal metastases, subtle lesions are easily missed, and the extent of disease is frequently underestimated compared with what a surgeon actually finds. Part of the difficulty is structural: the surgical PCI system, developed by Sugarbaker, was never designed for cross-sectional imaging, so radiologists assessing the abdomen on CT have lacked standardized regional boundaries. The result is inconsistent reporting and considerable variability between readers, which may explain why only a minority of radiologists currently attempt PCI scoring on images at all.

A recent Delphi consensus study involving 88 international experts in radiology, surgery, and gynecology changed that landscape by defining thirteen radiological PCI regions, numbered 0 through 12: the central region, right and left upper abdomen, epigastrium, left and right flanks, left and right lower abdomen, pelvis, and four small-bowel regions covering the upper and lower jejunum and ileum. The new research takes those expert-defined boundaries and teaches an artificial intelligence to reproduce them automatically. Crucially, the regions are not organ segmentations; they are subdivisions of the peritoneal cavity itself, each bounded by anatomical landmarks and surrounding organ surfaces—a spatial partition that is far harder for a neural network to learn than, say, the outline of a liver.

To train and test their models, the team assembled 62 contrast-enhanced CT scans from patients with gastric, ovarian, and colorectal cancers, deliberately spanning the full PCI range from 0 to 39. Every scan was painstakingly annotated in three dimensions by clinical researchers, with each annotation reviewed by a second annotator and disagreements settled by an expert radiologist. The effort involved was enormous: annotating a single scan takes roughly five hours. The researchers then compared two of the most widely used architectures in 3D medical image segmentation—the self-configuring convolutional network nnU-Net and the transformer-based Swin UNETR—using fivefold cross-validation and standard metrics including the Dice similarity coefficient, the 95th-percentile Hausdorff distance, and average surface distance.

The head-to-head comparison delivered a clear verdict. The convolutional nnU-Net achieved an overall Dice score of 0.81, outperforming the transformer-based Swin UNETR, which reached 0.76. Beyond raw accuracy, nnU-Net proved markedly more stable across patients, with a higher median Dice in every single region, whereas the transformer showed greater sensitivity to patient-specific anatomical variability. Both models struggled most in the right flank and the small-bowel regions, where regional boundaries are defined geometrically rather than by visible tissue contrast—precisely the places where a network trained purely on image appearance has nothing to latch onto.

That observation inspired the study’s most inventive contribution: an anatomically constrained pipeline that sidesteps the need to learn ill-defined boundaries at all. Instead of asking the network to distinguish, for example, the upper jejunum from the lower jejunum on a CT image where no visible border exists, the researchers trained a second nnU-Net on merged super-regions—grouping the flank and lower regions on each side, and all four small-bowel regions together. After prediction, a deterministic post-processing step splits these super-regions back into the thirteen individual regions using anatomical landmarks detected by the established TotalSegmentator tool. A transverse plane at the top of each hip bone divides flank from lower abdomen, while anteroposterior planes radiating from the ligament of Treitz—the duodenum–small-bowel transition—carve the small-bowel super-region into four equal-volume wedges, mirroring the consensus definitions.

The results were striking. The anatomically constrained pipeline lifted the overall Dice score from 0.81 to 0.84 and reduced boundary errors substantially, with the 95th-percentile Hausdorff distance falling from 13.7 to 11.8 millimeters and average surface distance from 4.1 to 3.4 millimeters. The largest gains came exactly where the baseline had faltered: in the small-bowel regions, mean Dice improved from 0.74 to 0.82, and average surface distance dropped from 5.8 to 3.7 millimeters. Residual errors in these regions were traced mainly to the localization of the anatomical landmarks themselves rather than to the segmentation network, pointing to a clear target for future refinement.

Perhaps the most clinically meaningful comparison in the study is not model against model but machine against human. On a separate subset of ten CT scans, each independently annotated in triplicate by three clinical researchers, the interobserver Dice score—the agreement between human experts—was 0.87. The automated pipeline, evaluated on identical data, achieved 0.84, with boundary-error metrics of the same order of magnitude as human variability. In most regions, the model’s agreement with human annotators approached the level at which humans agree with one another, and in the small-bowel regions the model actually showed far less variation across patients than the human observers did. In other words, the algorithm is already performing near the ceiling set by human consistency itself.

The clinical implications extend beyond a single score. Because current radiological assessment of peritoneal disease lacks standardized regional boundaries, an automatic segmentation overlay could guide radiologists to systematically inspect each predefined region, reduce interobserver variability, standardize reporting, and support surgical planning. It could also accelerate reporting workflows and serve as a training aid for less experienced readers. Computationally, the approach is feasible for real-world use: nnU-Net inference takes roughly seven seconds per scan on a single NVIDIA A100 GPU. The researchers envision a human-in-the-loop workflow in which the model proposes regions and radiologists make minor boundary adjustments, and a longer-term extension in which lesions are localized within each region and mapped to PCI subscores automatically.

The authors are candid about the limitations. The study is single-center, with a private dataset of 62 scans, and variation in scanners, acquisition protocols, and contrast timing across institutions could affect generalizability; external validation on multi-center data is a stated priority. The interobserver analysis rested on only ten scans, because triple-annotating the full cohort at five hours per scan was infeasible. Some patients with surgically altered anatomy or large tumors that displace organs produced outlier errors, particularly affecting the rule-based post-processing in the right flank. Yet the direction of travel is unmistakable. With code and trained model weights released openly on GitHub, and the radiological PCI framework itself born from international consensus, the study offers a reproducible foundation for something oncology has wanted for decades: an objective, noninvasive, and standardized measure of peritoneal cancer burden that any hospital with a CT scanner could one day compute in seconds.

Subject of Research: Deep learning segmentation of radiological peritoneal cancer index regions on CT for noninvasive staging of peritoneal metastases

Article Title: Deep learning-based segmentation of peritoneal cancer index regions from CT imaging

Article References: Gort, P. C., Fleurkens-Ewals, L. J. S., Kampmeijer, L. D., van Herwijnen, A. F., Tops-Welten, M. W., Claessens, C. H. B., Nederend, J., De Hingh, I. H. J. T., Lahaye, M. J., Luyer, M. D. P., & van der Sommen, F. (2026). Deep learning-based segmentation of peritoneal cancer index regions from CT imaging. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03796-9

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03796-9

Keywords: peritoneal metastases, peritoneal cancer index, deep learning, CT imaging, medical image segmentation, nnU-Net, Swin UNETR, TotalSegmentator, radiology, surgical oncology, artificial intelligence, computer-assisted surgery