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, led by Janine Rothert and colleagues spanning the University Hospital Schleswig-Holstein in Kiel, Otto-von-Guericke-University Magdeburg, and the University Medical Center Mainz, addresses a concept known as parenchyma-sparing hepatectomy. Unlike major resections that remove entire liver segments or lobes, parenchyma-sparing surgery minimizes the volume of liver tissue removed while maintaining oncological safety. This approach expands eligibility for curative surgery, reduces operative stress and morbidity, and preserves liver function for potential future resections. The stakes are quantifiable: clinicians must preserve a future liver remnant of at least 20 to 30 percent in healthy livers, and at least 40 percent in patients with impaired hepatic reserve. Accurate preoperative planning therefore has to integrate tumor location, arterial and portal venous perfusion, hepatic venous drainage, and biliary anatomy all at once.

To test how different levels of anatomical and functional information shape surgical predictions, the team built three distinct modeling paradigms, each operating on segmentation-derived representations of the liver, tumor, and vessels. The first is a purely geometric, distance-based approach designed for non-anatomical, surface-near tumor resections such as atypical wedge procedures. It uses an Euclidean distance transform to find the shortest path from tumor to liver surface, generates a global resection direction from averaged normalized direction vectors, and then dilates the tumor mask by an 8-millimeter margin, a figure matching the dataset’s average planned minimal resection margin of 7.63 millimeters, before extending the zone toward the liver surface. The result is a cylindrical, maximally tissue-sparing resection proposal that requires no knowledge of blood vessels at all.

The second paradigm is explicitly perfusion-aware. Starting from the same 8-millimeter dilation, the algorithm converts skeletonized portal and hepatic vein segmentations into directed graphs, placing nodes at bifurcations and inferring flow direction from the largest-radius node. Any vessel intersecting the initial resection zone is marked as transected, and all downstream branches are classified as non-perfused. Each liver voxel is then assigned to its nearest perfused or non-perfused vessel following a classical method from Selle and colleagues, defining the viable remnant and the final predicted resection zone. In effect, the model approximates Couinaud segments and simulates the physiological consequence of cutting a specific vessel, something purely geometric tools cannot do.

The third paradigm abandons hand-crafted rules entirely. The researchers formulated resection zone prediction as a volumetric segmentation task and trained a deep learning model on the nnU-Net framework, the self-configuring architecture that has repeatedly set benchmarks in medical image segmentation. Input consisted of three channels, a liver mask, a tumor mask, and a combined portal and hepatic vein mask, with the resection zone as the target. Training used 3D full-resolution patches of 160 by 192 by 224 voxels, a residual large encoder with six down- and upsampling stages, and a combined Dice and cross-entropy loss, run over 1000 epochs with fivefold cross-validation balanced for tumor and resection types.

The evaluation drew on a clinical database from Mainz comprising 312 patients who underwent 338 open liver tumor resections performed by seven different surgeons. After strict inclusion and exclusion criteria based on data integrity, entity integrity, and procedure type, 104 patients with 104 retrospectively defined resection zones remained for analysis. All segmentation masks, generated semiautomatically with surgical planning software, were rasterized onto a common voxel grid with 0.4-millimeter isotropic resolution, keeping boundary discretization errors below one millimeter. Predictions were scored with overlap metrics, the Dice similarity coefficient and intersection over union, and distance metrics, including the 95th-percentile Hausdorff distance, maximum Hausdorff distance, and average surface distance.

The results reveal a striking trade-off. The distance-based model produced predictions with the smallest surface deviation, a 95th-percentile Hausdorff distance of 33.89 millimeters, but suffered from systematic undersegmentation, achieving a Dice score of only 58.18 percent. The perfusion-based method struck the best overall balance, with a Dice score of 67.41 percent and a Hausdorff distance of 37.92 millimeters, though it tended to overestimate the resection region because of its strict binary assumption that tissue downstream of a cut vessel is entirely lost, whereas real hepatic perfusion is highly interconnected. The deep learning model delivered the highest overlap accuracy at 76.31 percent, but with the largest distance errors, a 95th-percentile Hausdorff distance of 65.21 millimeters, reflecting localized boundary inaccuracies and occasional spatially distant false positives that disproportionately inflated the distance metrics.

Statistical analyses added crucial nuance. Pearson correlation tests showed that all three models performed significantly better, in terms of Dice overlap, for deeper tumors, likely because larger tumors extend further into the liver and reduce the relative impact of minor boundary deviations. Higher ratios of tumor volume to resection zone volume also correlated with improved performance across models. Subgroup comparisons using the Mann-Whitney U test found that atypical, non-anatomical wedge resections were consistently harder to predict than standardized segmental resections, underscoring how individualized parenchyma-sparing surgery is, and how challenging it is to replicate algorithmically. Notably, the deep learning model’s advantage was strongest for larger and more standardized resection volumes, while the rule-based approaches retained their edge in constrained, tissue-sparing scenarios.

The authors are careful to position these tools as decision support rather than replacement for surgical judgment. Each paradigm, they argue, functions as a modular building block suited to a different clinical situation: the distance-based model is well suited to maximal parenchyma-sparing resections, particularly for patients with limited functional reserve; the perfusion-based model is advantageous when tumors sit near major vessels where downstream perfusion loss must be anticipated; and the deep learning model offers the strongest raw predictive power where larger resections are planned. Because the rule-based methods expose their logic through adjustable parameters such as the resection margin, they also offer a transparency and clinical controllability that neural networks currently lack.

Looking forward, the team proposes hybrid strategies that fuse the paradigms, integrating perfusion-aware and distance-based reasoning into deep learning either during training or as a post-processing step, adding surface-distance-aware loss functions, clinical metadata such as tumor entity and body composition, and simple connected-component cleanup to remove false positives. Extending the framework to standard segment-based hepatectomies is also on the agenda. With the code publicly released on GitHub, the study offers surgeons something genuinely new: not a single algorithmic answer, but a menu of differently nuanced resection proposals for the same patient, each grounded in a different model of liver anatomy and physiology. In a field where every spared millimeter of functional liver tissue can determine whether a patient tolerates the operation, and whether future curative options remain open, that flexibility could prove transformative.

Subject of Research: Computational modeling of resection zone prediction for parenchyma-sparing liver surgery planning

Article Title: Resection zone prediction for parenchyma-sparing hepatectomy planning: a comparative study of three modeling paradigms

Article References: Rothert, J., Rakshit, J., Salz, J. L., Huettl, F., Ehses, V., Huber, T., Lang, H., Saalfeld, S., & Hille, G. (2026). Resection zone prediction for parenchyma-sparing hepatectomy planning: a comparative study of three modeling paradigms. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03798-7

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03798-7

Keywords: liver cancer, hepatectomy, surgical planning, deep learning, nnU-Net, perfusion modeling, resection zone prediction, parenchyma-sparing surgery, medical image segmentation, Hausdorff distance, Dice similarity coefficient, computational surgery

Pulmonary hypoplasia, or underdeveloped lungs, is a final common pathway shared by many fetal diseases. Congenital diaphragmatic hernia, in which abdominal organs push into the chest, is the best-known example, but kidney malformations with low amniotic fluid, bladder outlet obstruction, bronchopulmonary malformations, and skeletal dysplasias can all leave a fetus with lungs too small to sustain life after birth. Fetal magnetic resonance imaging provides a way to quantify the problem: by tracing the lungs slice by slice, radiologists can calculate total lung volume and compare it with expected values for gestational age, producing numbers that guide decisions about delivery planning and even fetal surgery.

The bottleneck has always been the tracing itself. Manual segmentation of the lungs on fetal MRI typically takes more than thirty minutes, and longer when the anatomy is distorted, the fetus moves during the scan, or amniotic fluid levels are abnormal. Tertiary fetal imaging centers often perform only two to five fetal MRI examinations per day, partly because of this labor requirement. Existing automated models offered little help beyond the most common diagnosis, because most were trained exclusively on congenital diaphragmatic hernia cases or structurally normal fetuses, leaving the many other causes of pulmonary hypoplasia without a reliable tool.

To build a more broadly applicable system, the researchers assembled a retrospective dataset of fetal MRI examinations performed between January 2016 and March 2025. Of 1,203 examinations identified, 906 had complete, recoverable expert-adjudicated lung segmentations that had been created during routine clinical care and used for actual patient counseling. The cohort spanned gestational ages from 16 to 38 weeks, with a mean of 27.1 weeks, and included the full spectrum of pulmonary hypoplasia etiologies. The data were split into 724 training and 182 independent test examinations for total lung volume, with a smaller subset of 196 examinations, including thoracic cavity and mediastinal labels, used for a second model predicting percent predicted lung volume.

Technically, the team used nnU-Net v2, a self-configuring segmentation framework that automatically determines preprocessing steps such as image resampling, intensity normalization, and patch size. Models were trained on an NVIDIA A100 graphics processing unit using five-fold cross-validation with both two-dimensional and three-dimensional full-resolution configurations, and final predictions combined the two model types through an ensemble. Notably, the researchers deliberately worked with the clinically acquired coronal T2-weighted single-shot fast spin-echo image stacks rather than reconstructed three-dimensional volumes. These source stacks are available at the time of interpretation, require no extra reconstruction step, and are less dependent on slice-to-volume reconstruction, a design choice intended to support rapid, pragmatic deployment in real clinics.

The results were striking for total lung volume. The automated model achieved a mean Dice similarity coefficient of 0.81, indicating strong spatial overlap with expert contours, and an intraclass correlation coefficient of 0.955 with manual volumetry, a level conventionally classified as excellent. Bland-Altman analysis showed minimal systematic bias of about one milliliter, with 95 percent limits of agreement between roughly minus 15 and plus 13 milliliters. Median absolute percentage error was 11.1 percent, close to the ten percent benchmark the authors selected as approximating inter-reader variability in fetal lung volumetry. Average inference time was 3.8 seconds per examination, roughly a 450-fold speedup over manual tracing.

The second model, which estimates percent predicted lung volume by additionally segmenting the mediastinum and thoracic cavity, performed less well. Its Dice coefficient averaged 0.76, its intraclass correlation coefficient was 0.727, and median absolute percentage error rose to 14.5 percent. The authors attribute this to the compounded difficulty of segmenting multiple structures that are variably distorted by herniated abdominal contents, a thymus that resembles hypoplastic lungs on T2-weighted images, and mass effects that shift the chest anatomy. At the clinically meaningful level of risk-category assignment, automated observed-to-expected total lung volume matched manual assignments in 86.8 percent of test examinations, with most disagreements occurring between adjacent severity categories, whereas percent predicted lung volume agreed in only 54.1 percent of cases.

Performance was remarkably stable across the disease spectrum. The total lung volume model performed best on skeletal dysplasia, left-sided congenital diaphragmatic hernia, and renal disease, and remained reliable in the presence of severe motion artifact, oligohydramnios, and polyhydramnios. It struggled most with very large bronchopulmonary malformations whose lesion volume exceeded that of the lungs themselves, where the model sometimes mistook the lesion for lung tissue and overestimated lung volume. Accuracy also dipped in fetuses scanned before 20 weeks of gestation, though it improved with advancing gestational age, precisely when lung volume metrics matter most for predicting neonatal outcomes and planning interventions.

Crucially, the system is designed for human oversight rather than autonomous use. In a simulated human-in-the-loop workflow, an attending fetal radiologist reviewed and corrected the 25 lowest-performing outputs, those with a Dice coefficient below 0.70, in an average of just 3.1 minutes per case. Most corrections involved removing labels erroneously assigned to bronchopulmonary malformations or adjusting boundaries along the diaphragm and herniated contents. Combining four-second automated segmentation with a few minutes of expert adjudication still cuts the total workflow time roughly tenfold, potentially allowing fetal lung volumetry to be delivered on the same day as imaging, when counseling and multidisciplinary decisions are made.

The implications extend beyond the clinic. Scalable, reproducible lung segmentations are a prerequisite for radiomics-based research, in which small boundary inaccuracies can propagate into unreliable quantitative features across large heterogeneous datasets. The authors caution that their study was retrospective and single-institution, that risk-category thresholds were derived from left-sided congenital diaphragmatic hernia outcome data and may not generalize to other etiologies, and that external validation across institutions, scanners, and vendors remains necessary. Still, by collapsing a thirty-minute expert task into seconds while preserving expert review for the hardest cases, the work represents a concrete step toward making quantitative fetal lung assessment a routine, same-day part of prenatal care.

Subject of Research: Automated deep learning segmentation of fetal MRI lungs for rapid lung volume estimation in pulmonary hypoplasia

Article Title: Automated fetal magnetic resonance imaging lung segmentation for rapid lung volume estimation across diverse pulmonary hypoplasia phenotypes

Article References: Choi, J. J., Hegde, S., Dudley, J., Galganski, L. A., Kline-Fath, B. M., Somasundaram, E., & Dillman, J. R. (2026). Automated fetal magnetic resonance imaging lung segmentation for rapid lung volume estimation across diverse pulmonary hypoplasia phenotypes. Pediatric Radiology. https://doi.org/10.1007/s00247-026-06765-1

Image Credits: AI Generated

DOI: 10.1007/s00247-026-06765-1

Keywords: fetal MRI, lung segmentation, deep learning, pulmonary hypoplasia, congenital diaphragmatic hernia, nnU-Net, lung volume, prenatal diagnosis, artificial intelligence, medical imaging, pediatric radiology, human-in-the-loop

The research, led by Wei Tang and colleagues at the Affiliated Yongchuan Hospital of Chongqing Medical University, set out to do something deceptively simple but technically demanding: automatically quantify paraspinal muscle at the L3-L4 level of the lumbar spine on standard magnetic resonance images, without any manual slice selection and without any manual segmentation. The L3-L4 level is the conventional landmark for muscle assessments of the trunk, chosen because it sits at the center of the lumbar region and corresponds to the same plane used in computed tomography studies of body composition. Any automated system that hopes to be clinically useful must therefore find this exact level on its own, segment the muscle tissue accurately, and produce measurements that a clinician can trust.

To test whether their workflow could deliver, the team turned to the Chongqing Osteoporosis Screening Study, an external cohort quite separate from any data used to build the algorithm. From 152 screened participants, 146 eligible cases entered the automated pipeline. The results were striking on the first metric alone: 144 of the 146 cases were processed successfully, a success rate of 98.6 percent. In medical imaging automation, where a single misregistered scan can derail an entire analysis, near-total throughput on unseen external data is a meaningful achievement. It suggests the pipeline is robust to the everyday variability of clinical scans, from differences in patient anatomy to scanner settings.

The technical heart of the validation lies in three distinct performance domains. The first is localization, the task of automatically identifying the correct axial slice at L3-L4. In a validation subset of 30 cases selected by age-stratified random sampling, the automated localization was in essentially complete agreement with manual identification: the mean absolute slice error was just 0.07 slices, with a standard deviation of 0.26, and the hit rate was 100 percent. In plain terms, the algorithm never picked the wrong vertebral level, and even when it deviated at all, the deviation was a small fraction of a single slice. For a measurement that can shift noticeably if the imaging plane moves even one slice away from the target disc, this precision is foundational.

The second domain is segmentation, the pixel-by-pixel delineation of the paraspinal muscle regions of interest. Here the team evaluated overlap between the automated contours and a carefully reviewed manual reference, known as the final reviewed GT1 reference, using the Dice similarity coefficient, a standard measure that scores perfect overlap at 1.0. The automated segmentation achieved a Dice score of 0.934 with a standard deviation of only 0.016. That combination of a high mean and a very tight spread matters as much as the mean itself: it indicates the algorithm did not merely perform well on average but performed consistently across patients of different builds and ages. A Dice score above 0.93 is generally regarded as excellent in medical image segmentation, approaching the ceiling imposed by inter-observer variability between human experts.

The third and most clinically consequential domain is agreement of the derived measurements. The pipeline produces two key quantities: the lean cross-sectional area of the paraspinal muscles, which reflects the amount of contractile muscle tissue, and a fat fraction estimate derived from T2-weighted signal intensities, which serves as a proxy for fatty infiltration of the muscle. The authors are careful to note that this T2-based fat fraction estimate is not the same as proton density fat fraction, which would require chemical-shift imaging techniques; it is a signal-intensity-based approximation available from standard clinical sequences. Intraclass correlation coefficients between automated and manual measurements ranged from 0.91 to 0.98, with the lean cross-sectional area showing the more stable consistency. Values in this range indicate excellent agreement, the kind of concordance that regulatory frameworks for quantitative imaging biomarkers typically demand.

Robustness across age was a central concern, since muscle quantity and quality change dramatically over the lifespan and fatty infiltration increases with age, potentially confusing any signal-based measurement. The validation subset was deliberately age-stratified, and performance was broadly consistent across age groups, suggesting the pipeline does not silently degrade in older patients whose muscles are more fat-laden. This matters because sarcopenia and age-related muscle degeneration are precisely the populations in whom automated quantification could deliver the greatest clinical value, screening for muscle loss before it manifests as falls, disability, or poor surgical recovery.

The study is also notable for its candor about limits. In an exploratory analysis of a small degeneration subgroup of just six patients, the intraclass correlation coefficient for the T2-weighted fat fraction estimate carried a wide confidence interval, signaling substantial uncertainty in how well the automated estimate agrees with manual measurement when muscles are severely degenerated. The authors explicitly caution that agreement for this fat fraction estimate in degenerated muscle should be interpreted cautiously. This is an honest and important caveat: severe fatty replacement alters tissue signal in ways that may challenge intensity-based estimates, and a definitive answer will require larger degeneration-specific cohorts or sequences purpose-built for fat quantification.

Why does this matter beyond the radiology reading room? Paraspinal muscle atrophy and fatty infiltration have been linked in a growing body of literature to chronic low back pain, disc degeneration, and outcomes after spinal surgery. Yet because manual quantification is slow and expensive, muscle measurements rarely enter routine clinical decision-making. An automated workflow that runs on standard lumbar MRI, requires no manual slice selection, and achieves expert-level agreement changes that calculus entirely. Every patient already undergoing a lumbar MRI for back pain could receive, at no additional scanning cost, a quantitative muscle report: lean muscle area, fat fraction, and longitudinal trends across years of follow-up. Population-scale studies of muscle health, currently impractical, become feasible on existing image archives.

The study also illustrates a broader shift in how artificial intelligence tools for medicine are being evaluated. Rather than reporting performance only on internal test sets, where algorithms often flatter themselves, the Chongqing team performed a genuine external validation on an independent cohort, with age-stratified sampling and a manually reviewed reference standard. The pipeline’s reliance on the nnU-Net framework, a widely used self-configuring deep learning architecture for medical segmentation, suggests the approach could be adapted by other groups. The authors acknowledge the findings apply under the evaluated acquisition protocol, and wider multi-center, multi-scanner validation remains the natural next step. But the core message stands: the pieces needed to turn every spine MRI into an automatic muscle health assessment have now been demonstrated end to end, with success rates, localization accuracy, segmentation overlap, and measurement agreement all reaching levels that justify moving this technology from the laboratory toward the clinic.

Subject of Research: Automated quantification of paraspinal muscle at the L3-L4 level on lumbar MRI using deep learning segmentation with external validation

Article Title: Automated paraspinal muscle quantification at L3-L4 on lumbar MRI: external validation in a cohort study

Article References: Tang, W., Chen, T., Yan, Z., Yuan, M., Deng, M., & Shao, G. (2026). Automated paraspinal muscle quantification at L3-L4 on lumbar MRI: external validation in a cohort study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02868-1

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02868-1

Keywords: paraspinal muscles, lumbar MRI, automated segmentation, nnU-Net, external validation, lean cross-sectional area, fat fraction, deep learning, muscle atrophy, sarcopenia, medical imaging, BMC Medical Imaging

The research team, led by Patricia García-Berlanga and including Juan Zapata, Juan Martinez-Alajarin, and Pablo Hernandez-Cerdan, compared three distinct families of deep learning models for detecting stroke lesions on non-contrast MRI. Their motivation is straightforward: interpreting these scans currently depends heavily on the judgment of a radiologist, and automating even part of that process, as a support tool rather than a replacement, could accelerate triage and reduce the variability that naturally arises between different human readers. Building reliable models for FLAIR images has proven stubbornly difficult, however, precisely because of the features that make these scans attractive in the first place. The contrast between healthy and injured tissue is lower than in contrast-enhanced imaging, and the images can look markedly different depending on the scanner model or the hospital that produced them. Any automated system must therefore be robust not only to the subtle appearance of the lesions themselves but also to the considerable technical variation across imaging hardware.

To make the comparison meaningful, the researchers evaluated three well-known image-segmentation architectures under strictly matched conditions. The first was a classical three-dimensional U-Net, a manually tuned convolutional neural network that has served as a workhorse of medical image segmentation for years. The U-Net design uses an encoder that compresses an image into increasingly abstract feature maps and a decoder that expands those features back into a full-resolution prediction, with skip connections that preserve fine spatial detail. The second contender was nnU-Net, a newer self-configuring framework that automatically adapts its preprocessing, architecture, and training schedule to the specific dataset it is given, dramatically reducing the need for laborious manual trial and error. The third was an adapted version of YOLOv8, an object-detection architecture better known for real-time applications such as video analysis, which the authors modified to perform two-dimensional segmentation of MRI slices.

All three models were trained and tested on the ISLES 2022 dataset, a publicly available benchmark of 250 stroke MRI cases that is widely used in the lesion-segmentation community. Crucially, the same preprocessing pipeline and the same evaluation procedures were applied to every model, ensuring that differences in performance could be attributed to the architectures themselves rather than to unequal treatment. Performance was quantified with two standard metrics for segmentation accuracy: the Dice similarity coefficient and the Intersection over Union, commonly abbreviated IoU. Both metrics measure how closely the AI’s predicted lesion boundary overlaps the lesion as delineated by expert annotators, with a score of one indicating perfect agreement and a score of zero indicating no overlap at all. Because these metrics penalize both missed tissue and falsely labeled healthy tissue, they provide a balanced picture of where each model excels and where it falters.

The results delivered a clear verdict on accuracy. The self-configuring nnU-Net came out on top, achieving the highest average overlap with expert-marked lesions, with a Dice score of approximately 0.483. While that number may appear modest on an absolute scale, it reflects the genuine difficulty of the task on contrast-free images, where lesion boundaries are ambiguous even for human experts. Importantly, the nnU-Net held up comparatively well across the wide variety of lesion shapes and sizes represented in the dataset, suggesting that its automatic configuration strategy successfully generalizes to the heterogeneity that characterizes real clinical material. The framework’s ability to tune itself to a new dataset, rather than relying on hand-crafted settings, appears to be the decisive advantage in this setting.

The adapted YOLOv8 model was not far behind on precision, posting a Dice score of around 0.476, a difference small enough that the two leading approaches are functionally comparable in terms of overlap quality. What distinguished YOLOv8 was speed. It processed scans substantially faster than the other two methods, and the authors suggest that this could make it a strong candidate for time-sensitive clinical workflows where throughput matters as much as raw accuracy. In an acute stroke setting, where imaging findings must reach the treating physician within minutes, a model that sacrifices a sliver of precision for a large gain in processing time may ultimately deliver more clinical value than a slower but marginally more accurate alternative. This speed-accuracy trade-off is a recurring theme in deployed medical AI systems, and the study provides concrete numbers to inform that decision.

The classical 3D U-Net, included as a manually configured baseline, performed noticeably worse overall, with a Dice score of roughly 0.294. The researchers attribute this gap to the architecture’s greater sensitivity to manual parameter choices and to variation across the dataset. In other words, when the settings are chosen by hand rather than learned automatically, small suboptimal decisions can compound into substantially degraded performance, particularly when the training data span a broad range of lesion appearances and scanner characteristics. The result is a cautionary tale for research groups still relying on fully manual pipelines: the additional engineering effort of self-configuring frameworks appears to pay measurable dividends on challenging, heterogeneous clinical data.

Beyond the headline numbers, the study dug into the specific failure modes shared by all three approaches. Common trouble spots included small lesions located near the outer edges of the brain, where anatomical complexity and partial-volume effects confound the models. Diffuse, low-contrast regions of damage also proved inherently hard to distinguish from healthy tissue, since the signal differences that define such lesions sit close to the noise floor of the image. A third recurring problem was spatial misalignment, in which the predicted lesion outline did not correspond well to the actual location of the abnormality within the scan. The authors also identified preprocessing techniques that helped mitigate these weaknesses, including resampling to standardize the voxel spacing across scans, normalizing pixel intensity values so that images from different scanners share a common scale, and training the networks on smaller image patches rather than whole volumes at once, which increases the effective diversity of training examples and focuses learning on local texture.

From these findings, the researchers distilled practical recommendations for teams building similar tools. Where computing resources permit, self-configuring frameworks such as nnU-Net are worth prioritizing for their robustness and reduced dependence on expert tuning. In environments where processing speed is the binding constraint, lightweight architectures in the mold of the adapted YOLOv8 represent a reasonable and defensible trade-off. Regardless of which architecture is chosen, the authors stress that rigorous validation across multiple clinical sites and scanner types is essential before any such tool can responsibly be considered for real-world clinical use, given how profoundly scanner-to-scanner variability can affect performance. A model that excels on a single institutional dataset may falter on images from an unfamiliar machine, and only multi-site testing can reveal whether a system has genuinely learned the imaging signature of stroke or merely the quirks of one scanner population.

The study, titled Detection of Vascular Pathologies in Non-contrast Magnetic Resonance Images for Early Screening, contributes a carefully controlled benchmark that clinicians and engineers alike can use to calibrate expectations for automated stroke screening. Its central message is encouraging: modern self-configuring and lightweight deep learning models can now extract clinically meaningful lesion information from the fast, contrast-free MRI sequences that emergency departments already rely on, without the delays and costs of contrast administration. At the same time, the modest absolute Dice scores and the catalog of shared failure modes make clear that these systems remain support tools rather than autonomous diagnosticians. The path forward, the authors suggest, lies in combining the accuracy of self-configuring frameworks with the speed of streamlined detectors, validated across the messy diversity of real-world scanners, so that automated lesion detection can one day help shave precious minutes off the interval between a patient’s arrival and a life-saving treatment decision.

Subject of Research: Deep learning segmentation of stroke lesions on non-contrast FLAIR MRI scans

Article Title: AI models put to the test for spotting stroke damage on routine brain scans

Article References: AI models put to the test for spotting stroke damage on routine brain scans. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, deep learning, stroke, MRI, FLAIR, nnU-Net, YOLOv8, U-Net, image segmentation, ISLES 2022, medical imaging, clinical decision support

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

The study, registered retrospectively on ClinicalTrials.gov as NCT07162168, enrolled 4,152 participants who had undergone noncontrast routine CT scans. The cohort included 1,947 men, with a median age of 59 years spanning a remarkable range from 18 to 99. Rather than relying on slow, labor-intensive manual tracing of the femur, the team trained a deep learning segmentation model based on the nnU-Net framework, a self-configuring architecture widely regarded as one of the most robust tools for biomedical image segmentation. A subset of proximal femurs was first annotated by hand to teach the model exactly where the bone begins and ends in each volumetric scan. Once trained, the automated system could carve out the femoral volume of interest on its own, achieving a Dice similarity coefficient of 0.963 on the test set, a score indicating near-perfect overlap with expert annotations.

With the femur reliably isolated in thousands of scans, the researchers turned to radiomics, the practice of converting medical images into hundreds of quantitative features that describe texture, shape, intensity and spatial patterns within a defined region. Using the open-source PyRadiomics package, they extracted features from each segmented femur, following standardized conventions designed to make radiomic measurements reproducible across scanners and institutions. The idea behind radiomics is deceptively simple: the microscopic architecture of bone, including trabecular thinning, cortical porosity and the redistribution of mineral, leaves fingerprints in the statistical texture of a CT image. Those fingerprints, aggregated across thousands of features, may encode a person’s skeletal age far more richly than a single density measurement.

To test whether these features actually carry information about aging and sex, the team built machine learning models in a development cohort of 2,969 participants and then evaluated them in a completely separate external validation cohort of 1,183 participants, a design that guards against the optimistic results that plague many single-center AI studies. For age, the best performer was a support vector regression model, which estimated chronological age from femoral radiomics with a coefficient of determination of 0.50 in external validation, a mean absolute error of 8.84 years and a root mean squared error of 11.32 years. In plain terms, the model could roughly halve the uncertainty of a blind guess about a person’s age using nothing but the texture of their hip bone.

Sex classification proved even more striking. Two algorithms, support vector classification and the gradient-boosting method XGBoost, both achieved an area under the receiver operating characteristic curve of 0.94 in external validation, with support vector classification showing slightly higher accuracy and F1-score. An AUC of 0.94 means the model distinguished male from female skeletons with high reliability across the entire cohort, capturing structural sexual dimorphism in the femur that goes well beyond simple size differences. Because the features were extracted automatically from routine scans, the result hints that sex-related skeletal architecture is legible in ordinary clinical imaging without any dedicated bone protocol.

Perhaps the most consequential finding concerns the relationship between the radiomics-derived age index and actual bone mineral biology. In the external validation cohort, the researchers compared their radiomic age index against volumetric bone mineral density of the femoral neck measured by quantitative computed tomography, the gold-standard three-dimensional density technique. The radiomics index tracked vBMD more strongly than chronological age did, with a Spearman correlation of minus 0.546 versus minus 0.443 for age alone, a difference the authors confirmed statistically using Steiger’s test (Z = 4.97, P < 0.001). This comparison matters because it suggests the radiomic signature is not merely a proxy for how many birthdays a person has had, but a reflection of genuine structural deterioration in the bone, the kind of change that underlies osteoporosis and hip fracture risk.

The clinical logic here is what researchers call opportunistic imaging. Dual-energy X-ray absorptiometry, the standard test for osteoporosis, is underused: many people at risk never get scanned, and fractures often arrive as the first sign of disease. But a large fraction of older adults undergo abdominal, chest or pelvic CT for unrelated reasons, and those scans routinely capture the proximal femur. If an automated pipeline can convert that incidental imaging into a skeletal age index or a bone quality flag, clinicians could identify patients with accelerated bone aging who would benefit from confirmatory density testing or preventive treatment, all without a single additional scan, extra radiation dose or added visit. The authors caution that their radiomic patterns still need validation against hard clinical outcomes such as incident fractures before such screening becomes practice, but the biological anchoring to vBMD is an encouraging first step.

The technical achievement should not be undersold either. Automated segmentation at a Dice coefficient above 0.96 removes the bottleneck that has historically limited radiomics to small, single-center studies requiring hours of expert contouring. By combining nnU-Net segmentation, standardized PyRadiomics extraction and rigorous external validation across nationwide cohorts, the study offers a template for how imaging biomarkers can be scaled to population level. The work was supported by the Beijing Natural Science Foundation-Fengtai Joint Innovation Project, the National Natural Science Foundation of China Science Center Program and the Shenzhen Medical Research Fund, and it was approved by the ethics committee of Peking University People’s Hospital with informed consent waived for the retrospective use of de-identified images.

There are, of course, caveats to keep in mind. The cohort was drawn from Chinese hospitals, and radiomic features can be sensitive to scanner vendor, reconstruction settings and dose, so multicenter generalization beyond this dataset remains to be demonstrated. The age model’s mean absolute error of nearly nine years is impressive for image-only prediction but far too coarse to serve as a precise biological clock for any individual. And the study measured associations with density, not with fractures, falls or mortality, so the ultimate clinical value hinges on longitudinal follow-up. Still, the direction of travel is clear: the femur that appears incidentally on a routine scan carries a quantifiable structural biography, written in texture patterns that machines can now read at scale.

If subsequent studies confirm that radiomics-derived bone age predicts fractures and responds to osteoporosis treatment, the implications could extend well beyond orthopedics. Bone is a living archive of hormonal change, nutrition, inflammation and mechanical loading, and a cheap, automated readout of its microarchitecture could inform research on aging itself, from sarcopenia to metabolic bone disease. For now, the study stands as a proof of concept that the skeleton’s quiet remodeling, normally invisible in the reading room, can be decoded from data that hospitals already possess, waiting in the archive.

Subject of Research: Automated CT radiomics of the proximal femur for quantifying age- and sex-related bone structural changes

Article Title: Automated proximal femur CT radiomics identifies age- and sex-related radiomic patterns of bone structure: a nationwide multicohort validation study

Article References: Cheng, H., Zhuang, L., Zhang, Y., Ran, Y., Jiang, B., & Kou, Y. (2026). Automated proximal femur CT radiomics identifies age- and sex-related radiomic patterns of bone structure: a nationwide multicohort validation study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02836-9

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02836-9

Keywords: radiomics, proximal femur, CT imaging, deep learning, nnU-Net, bone aging, bone mineral density, osteoporosis, machine learning, opportunistic screening, quantitative CT, hip fracture risk

The research, led by Sanaa Amina Gourine of the Nara Institute of Science and Technology together with colleagues from Miyazaki University, Shiga University of Medical Science, Osaka University, and Ehime University, describes a pipeline for volumetric segmentation and quantitative analysis of 21 torso structures in CT images. Segmentation, in this context, means teaching an algorithm to trace the exact boundaries of each muscle and fat compartment voxel by voxel across a three-dimensional scan. Once those boundaries are known, software can compute each structure’s volume in cubic centimeters and its mean density in Hounsfield units, the standard radiological measure of tissue composition. Lower muscle density on CT reflects fatty infiltration of the muscle, a hallmark of aging and disuse that correlates with reduced strength and physical performance. Extracting these numbers automatically from thousands of existing scans would turn routine clinical imaging into an opportunistic screening resource for conditions that currently go undiagnosed.

The technical heart of the study is a comparison between two deep learning architectures, both augmented with a mechanism for estimating predictive uncertainty. The first is a two-dimensional Bayesian UNet, a convolutional neural network that processes the scan slice by slice and uses a technique called Monte Carlo dropout sampling. Instead of producing a single deterministic answer, the network runs its prediction multiple times with randomly deactivated internal units, and the variability across those runs serves as a proxy for how confident the model is at each location. The second, and ultimately the winning, approach is a fully three-dimensional, patch-based model built on the well-known nnU-Net framework, which the authors call the 3D MSKSegmenter. Rather than relying on repeated stochastic sampling, this model derives uncertainty directly from the entropy of its prediction probabilities: when the network’s output probabilities are spread evenly across competing classes, entropy is high and the prediction is flagged as unreliable.

That uncertainty estimate is not a mathematical flourish; it is the framework’s quality control system. In the validation experiments, the researchers found strong negative correlations between predictive uncertainty and segmentation accuracy, measured by the Dice coefficient, the standard overlap metric comparing automated contours against manual ground truth. The correlation reached minus 0.870 for the 3D MSKSegmenter and an even steeper minus 0.955 for the two-dimensional Bayesian UNet. In practical terms, when the model reports low uncertainty, its segmentation is almost always accurate, and when uncertainty climbs, accuracy tends to collapse. This relationship allows the framework to act as its own inspector: for large-scale database studies where no expert can review every output, the system can automatically filter out cases whose uncertainty exceeds a threshold, ensuring that only reliable measurements flow into downstream statistical analyses. The authors validated this quality control mechanism across a fivefold cross-validation experiment on 20 manually labeled CT images.

The 3D MSKSegmenter outperformed both its two-dimensional counterpart and a formidable external benchmark. The team compared their results against TotalSegmentator, a widely used, state-of-the-art open-source model capable of segmenting 104 anatomical structures across the whole body. Despite that model’s breadth, the torso-specialized 3D MSKSegmenter achieved a superior average Dice coefficient of 0.89 plus or minus 0.03 across the 21 target structures, a statistically significant advantage over both competitors at p less than 0.001. The result underscores a recurring lesson in medical image analysis: models trained specifically on the structures and imaging characteristics of a target anatomy, with careful handling of the enormous variability in body habitus, scan protocol, and contrast phase found in real hospital databases, can beat generalist systems on their home turf.

Robustness was then tested at scale. The researchers applied the framework to a single-institution database of approximately 1,200 patients, an external test set far larger than the labeled training and validation data. The framework proved stable across this heterogeneous clinical population, and the uncertainty-based filtering continued to identify the cases most likely to contain segmentation errors. This matters because the ultimate ambition of the project is not to segment 1,200 scans but to enable population-scale studies in which tens of thousands of torso CTs, already acquired for other reasons, are mined for musculoskeletal biomarkers without any additional imaging cost or radiation exposure. Opportunistic screening of this kind has been proposed for cardiometabolic risk and sarcopenia, but it has been limited by the absence of segmentation tools that are simultaneously accurate across many muscle groups and honest about their own failure modes.

The downstream analysis delivered the kind of biological signal that justifies the engineering effort. Measuring muscle volume and mean Hounsfield units across the large database, the researchers detected significant age- and sex-related trends in muscle density and volume, consistent with the known epidemiology of sarcopenia and age-related fatty muscle degeneration. Because the framework quantifies 19 individual muscles rather than a single aggregate muscle area at one reference slice, it opens the door to site-specific analyses: whether the gluteal muscles, the paraspinal stabilizers, or the abdominal wall degrade at different rates, and how those patterns relate to conditions such as hip osteoarthritis, knee osteoarthritis, or postoperative physical function. Prior work by overlapping research groups has linked gluteal muscle volume and quality to early recovery after total hip arthroplasty and lower limb muscle degeneration to knee osteoarthritis severity, and a scalable torso-wide tool provides the missing population-level instrument for such investigations.

The study also situates itself within a rapidly maturing field. Earlier approaches to automated muscle segmentation relied on random forest classifiers, atlas-based registration, or fuzzy c-means clustering, typically targeting one or a few muscle groups. The rise of deep learning brought UNet-style architectures and, more recently, the self-configuring nnU-Net, which has become a de facto standard in biomedical segmentation. Bayesian extensions of the UNet introduced calibrated confidence estimation to medical imaging, and previous validations demonstrated that uncertainty estimation works for bone and muscle assessment in hip-to-knee CT. The present work extends that lineage in two directions: it covers an unusually comprehensive set of torso musculature and fat compartments, and it explicitly integrates uncertainty as a gatekeeper for biomarker extraction in very large databases, rather than treating it as a laboratory curiosity.

Transparency is part of the package. The code for the 3D MSKSegmenter, built on the open nnU-Net repository, has been released on GitHub by the NAIST Bio-Image Intelligence group, and the trained models are available from the corresponding authors on reasonable request. The retrospective study was approved by local ethics committees and conducted in accordance with the Helsinki Declaration, with informed consent obtained through opt-out mechanisms. Funding came from the Japan Society for the Promotion of Science, the Japan Agency for Medical Research and Development, and the Cross-ministerial Strategic Innovation Promotion Program on integrated health care systems.

The broader implication is that the raw material for a new generation of musculoskeletal epidemiology is already sitting on hospital servers. As populations age, sarcopenia and related muscle disorders are projected to become a major driver of disability and health care cost, yet muscle status is rarely measured outside specialized research settings. A framework that can sweep through existing CT archives, quantify 21 body composition structures per patient, and flag its own uncertain outputs for review turns routine diagnostic imaging into a longitudinal window on the aging musculoskeletal system. If external validations at additional institutions replicate the accuracy and reliability reported here, the humble torso CT, acquired for everything from lung cancer screening to abdominal pain, could quietly become one of the most valuable population health instruments in modern medicine.

Subject of Research: Automated deep learning segmentation and uncertainty-based quality control for volumetric musculoskeletal analysis of torso CT images

Article Title: Automated framework for volumetric musculoskeletal segmentation and assessment of torso CT images in large-scale databases

Article References: Gourine, S. A., Soufi, M., Otake, Y., Masaki, Y., Murakami, Y., Nagatani, Y., Watanabe, Y., Uemura, K., Kono, K., Takao, M., Fujimori, T., Sugano, N., & Sato, Y. (2026). Automated framework for volumetric musculoskeletal segmentation and assessment of torso CT images in large-scale databases. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03785-y

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03785-y

Keywords: musculoskeletal segmentation, torso CT, deep learning, nnU-Net, predictive uncertainty, Bayesian UNet, sarcopenia, body composition, TotalSegmentator, Hounsfield units, large-scale databases, opportunistic screening

The study, led by Hye Ju Yeo and Ha Lim Kim of Pusan National University Yangsan Hospital in South Korea, together with colleagues including Woo Hyun Cho, examined 1,107 adults with sepsis who had undergone abdominal CT imaging as part of their clinical care. Rather than relying on height and weight, the team deployed an automated segmentation pipeline built on the nnU-Net framework, a self-configuring deep learning architecture that has become a standard tool for biomedical image analysis, supplemented by the TotalSegmentator software capable of delineating more than one hundred anatomic structures in a single scan. The algorithms measured the cross-sectional areas of subcutaneous adipose tissue, the fat layer beneath the skin, and visceral adipose tissue, the fat packed around internal organs, at the level of the third and fourth lumbar vertebrae, a standard landmark in body composition research.

The technical rationale for this approach is straightforward. Body mass index cannot distinguish between fat stored under the skin and fat stored deep within the abdomen, yet these two depots behave in profoundly different ways biologically. Subcutaneous adipose tissue acts largely as a passive energy reservoir and, in some contexts, appears metabolically protective, sequestering lipids away from the liver, muscle, and bloodstream. Visceral adipose tissue, by contrast, drains directly into the portal circulation, is richly innervated and hormonally active, and is strongly associated with insulin resistance, systemic inflammation, and adverse outcomes across a range of acute and chronic illnesses. The idea that these depots might carry opposite prognostic signals during sepsis is not new, but measuring them reliably has traditionally required painstaking manual tracing of CT images, which is impractical at scale.

Among the 1,107 patients in the cohort, 300, or 27.1 percent, died within 28 days of their sepsis diagnosis during the index hospitalization. When the researchers fed their imaging-derived measurements into Cox proportional hazards models adjusted for clinically selected covariates, a clear and internally consistent picture emerged. Greater subcutaneous adipose tissue area was associated with lower mortality, with an adjusted hazard ratio of 0.83 per one standard deviation increase, a statistically significant protective association. Obesity itself, defined here as a body mass index of 25 kilograms per square meter or higher in keeping with thresholds commonly used in Asian populations, was also associated with lower mortality, with an adjusted hazard ratio of 0.62. But the most telling variable was the ratio of visceral to subcutaneous fat: patients with a higher VAT/SAT ratio had significantly higher mortality, with an adjusted hazard ratio of 1.14 per standard deviation increase.

That pattern reframes the obesity paradox in mechanistic terms. It suggests that what has appeared to be a survival advantage of obesity in sepsis may actually reflect the composition of body fat rather than its total quantity. A patient whose excess weight is dominated by subcutaneous fat may be metabolically better buffered against the catabolic storm of critical illness than a leaner patient whose fat is disproportionately visceral. The finding aligns with a growing body of work. A 2016 study in Critical Care Medicine by Pisitsak and colleagues found that an increased visceral-to-subcutaneous adipose tissue ratio in septic patients was associated with adverse outcomes, and subsequent research suggested that the survival benefit of a low ratio depends on the balance of LDL cholesterol clearance versus production during infection. Population-level analyses in the UK Biobank have similarly shown that abdominal fat distribution outperforms traditional anthropometric indices in predicting sepsis outcomes.

The stratified analyses added further texture, though with appropriate statistical caution. In exploratory models stratified by body mass index category, the protective association of subcutaneous fat was most pronounced among underweight patients, where greater subcutaneous adipose tissue area carried an adjusted hazard ratio of 0.37, meaning a substantially lower hazard of death. Conversely, among patients with obesity, a higher visceral-to-subcutaneous ratio was associated with a markedly elevated mortality risk, with an adjusted hazard ratio of 1.48. In other words, not all obesity is created equal: a patient with obesity whose fat is predominantly subcutaneous may carry a very different prognosis from one whose obesity is visceral-dominant. The authors were careful to note, however, that formal interaction tests between these imaging measures and body mass index category did not reach statistical significance, with P values of 0.145 and 0.127, meaning the apparent effect modification could reflect chance and requires confirmation in larger, independent cohorts.

The study also interrogated skeletal muscle, a dimension of body composition that has gained increasing attention in critical care. Sarcopenia, the loss of muscle mass and function, is common in critically ill patients and has been linked to worse long-term outcomes, partly because sepsis itself induces profound muscle wasting through disrupted autophagy and mitochondrial dysfunction. The team quantified psoas muscle measures at the same lumbar level, including the psoas muscle index, a conventional marker of muscle quantity. Strikingly, muscle size itself did not retain an independent association with mortality after adjustment. What did survive the statistical scrutiny was muscle quality: a higher proportion of low-attenuation muscle, which reflects fatty infiltration of the muscle and is a radiographic signature of myosteatosis, was associated with higher mortality, with an adjusted hazard ratio of 1.14. A joint phenotype combining a high visceral-to-subcutaneous ratio with a low psoas muscle index, which the investigators had hypothesized might identify particularly high-risk patients, was not significant after adjustment.

These findings carry real implications for how risk is characterized in the intensive care unit. Abdominal CT scans are already obtained routinely in many sepsis patients to search for the infectious source, meaning the raw imaging data needed for automated body composition analysis often exists before anyone thinks to use it prognostically. A pretrained deep learning pipeline can extract these measurements in seconds without manual labor, and the underlying software is publicly available, making the approach reproducible and potentially deployable as a clinical decision-support tool. If validated externally, a VAT/SAT ratio or a low-attenuation muscle fraction could complement, and in some cases correct, the crude signal provided by body mass index, helping clinicians identify high-risk patients who would be missed by conventional anthropometry, including visceral-dominant patients with obesity and sarcopenic patients of normal weight.

The authors and independent commentators are quick to emphasize the study’s limitations. It was a single-center, retrospective cohort, which raises the possibility of selection bias, since only patients who happened to receive abdominal CT imaging could be included. The subgroup findings, however compelling, rest on exploratory analyses with non-significant interaction tests, and the skeletal muscle results in particular require external validation before they can inform practice. Residual confounding by illness severity, nutritional status, and chronic disease cannot be excluded in any observational design. Still, the convergence of this work with prior mechanistic and epidemiological evidence, and the elegance of using artificial intelligence to turn diagnostic scans that already exist into prognostic information that costs nothing extra, marks a meaningful step forward. The obesity paradox in sepsis has long been an embarrassment to simple models of body size and disease; deep learning analysis of where the body stores its fat may finally explain what the scale never could.

Subject of Research: Deep learning–based CT analysis of abdominal fat distribution and its association with 28-day mortality in sepsis patients

Article Title: Deep learning–derived abdominal adiposity phenotypes and 28-day mortality in sepsis

Article References: Yeo, H. J., Kim, H. L., Kim, K., Jang, J. H., Choi, E., Seol, H. Y., Lee, S. E., & Cho, W. H. (2026). Deep learning–derived abdominal adiposity phenotypes and 28-day mortality in sepsis. International Journal of Obesity. https://doi.org/10.1038/s41366-026-02220-1

Image Credits: AI Generated

DOI: 10.1038/s41366-026-02220-1

Keywords: sepsis, obesity paradox, visceral adipose tissue, subcutaneous adipose tissue, deep learning, nnU-Net, body composition, computed tomography, 28-day mortality, psoas muscle, myosteatosis, critical care