The research team, led by surgeons and engineers from institutions across India and the United Kingdom, built a system called Hernia Image Assist, or HIA. The team trained TransUNet-inspired models on 215 abdominal CT volumes from adults with ventral hernias, including umbilical, epigastric, incisional, and lateral types. Expert annotations by radiologists with more than five years of abdominal imaging experience served as ground truth. The automated segmentation achieved Intersection over Union values of 0.85 for hernia defects, 0.89 for rectus abdominis muscles, 0.87 for lateral abdominal wall muscles, and 0.91 for psoas muscles, with hernia sacs at 0.82. Converted to Dice similarity coefficients, these correspond to values between 0.90 and 0.95, a range consistent with deep learning benchmarks for muscle segmentation at the L3 vertebral level.

The clinical motivation is grounded in a well-documented measurement problem. Studies cited by the authors report that variability in defect measurement affects surgical decisions, including mesh selection, operative approach, and the need for component separation, in up to 56 percent of cases. Intraclass correlation coefficients for manual measurements across different modalities range from 0.21 to 0.737. Disagreement between radiologists and surgeons on CT interpretations occurs in 73 percent of cases, and experienced abdominal wall reconstruction surgeons correctly identify previous repair types in fewer than 50 percent of cases. These inconsistencies directly influence operative decision-making in complex presentations such as large defects, loss of domain, and multiply recurrent hernias.

The technical architecture behind HIA reflects the specific demands of hernia imaging. The pipeline begins with a preprocessing stage in which a deep learning model identifies the abdominal region in the input CT volume. Once the abdominal region is confirmed, the system deploys multiple segmentation models and computer vision algorithms to generate clinical metrics, including defect width, European Hernia Society classification for defect localization, the Tanaka index, and volumetric ratios comparing hernia sac volume to abdominal cavity volume. The L3 and psoas detection models were developed to measure Hounsfield unit attenuation, enabling prediction of sarcopenia, the age-related loss of muscle mass that influences surgical outcomes.

Each segmentation model shares a common TransUNet-inspired architecture, a hybrid design that combines a convolutional encoder with a vision transformer. The transformer processes patches of size 4 by 4, allowing the model to capture long-range spatial relationships that pure convolutional networks can miss. For models taking 128 by 128 inputs, the team removed two encoder and decoder layers to fit the smaller resolution. Training used the TensorFlow framework on an RTX 4090 24 GB GPU with a batch size of 10, the Adam optimizer, and binary cross-entropy loss. Models were trained for different numbers of epochs depending on the task: 600 for abdominal detection, 360 for defect segmentation, 540 for sac segmentation, and 670 for rectus segmentation. Data were split 80 percent for training and 20 percent for validation, and the training and validation loss curves ran in parallel, indicating minimal overfitting.

The annotation strategy was deliberately unconventional. Rather than labeling every structure on every scan, the team allocated non-overlapping subsets: 85 CTs for annotating the xiphoid process and pubic symphysis, another 85 for the hernia defect and sac, and 45 for the abdominal cavity, rectus muscles, and psoas muscles. Structures were annotated in the imaging plane that provided optimal visibility, with sagittal slices for the xiphoid and pubic symphysis, coronal slices for vertebral landmarks from the public VerSe20 dataset, and 512 by 512 axial slices for the hernia and muscular structures. The authors argue that annotating all structures simultaneously on every scan would have introduced annotation fatigue and reduced labeling consistency, a recognized risk in large-scale medical image labeling. The sparse annotation approach, with roughly one to three representative slices per structure per scan, is consistent with weakly-supervised and semi-supervised segmentation practices where dense volumetric labeling is resource-prohibitive.

Preprocessing applied a Hounsfield unit window of minus 200 to 200 to all CT volumes, followed by resizing and min-max normalization. The scans were supine breath-hold acquisitions of the abdomen and pelvis without intravenous contrast, collected on multidetector 8- to 64-slice scanners across two Indian institutions after institutional review board approvals. The algorithm was designed to accommodate wide variations in tube current and slice thickness. Surgical mesh was annotated on the same 85-scan subset used for defect and sac labels, with 10 to 20 axial slices per scan, allowing the model to learn the spatial relationship between mesh and defect directly. Mesh annotation posed particular challenges: low contrast between the mesh material and surrounding soft tissue made precise boundary delineation difficult, and proximity to the hernia defect sometimes obscured mesh edges.

The time savings are a central part of the claim. The manual annotation process in this study required 25 to 30 minutes per case, reflecting the time-intensive nature of comprehensive hernia evaluation. The automated system processes the same parameters within a couple of minutes. The authors position this efficiency alongside consistency: automated systems offer reproducible measurements by design, addressing the inter-observer variability that complicates both clinical care and outcomes research. They note that volumetric CT analysis, including hernia sac and abdominal cavity volume calculation, has been proposed as a comprehensive approach to complex hernia assessment, and their system attempts the same computation.

The study advances beyond earlier automated hernia segmentation work in scale and scope. Foundational studies by Xu and colleagues demonstrated abdominal wall delineation in ventral hernia CT using level-set methods in four cases, achieving mean surface errors under 2 millimeters, and later extended this to 20 cases with texture-augmented level-set segmentation. The present work deploys multi-structure deep learning segmentation across 215 cases and five anatomical structures. For context, other AI systems in hernia surgery have focused on outcome prediction rather than anatomical segmentation: an Atrium Health neural network reports 75 percent accuracy for predicting component separation requirements, and machine learning models from other institutions demonstrate 84 to 85 percent accuracy for recurrence and readmission prediction. The Elhage group’s convolutional neural network, trained on 9,303 CT images from 369 patients, achieved 81.3 percent accuracy in surgical complexity prediction.

The authors are explicit that this is a proof of concept, not a deployable clinical tool. All reported IoU values were derived from a held-out validation split of the same institutional dataset, with no independent external test set. Formal inter-rater agreement statistics between the two annotating radiologists were not prospectively computed, and metrics beyond IoU and Dice, including precision and recall, were not reported. No subgroup analysis by hernia type or defect size was performed, and the downstream clinical metrics such as defect width and volumetric ratios were not formally validated against clinical measurements. The training dataset’s geographic limitation to two Indian institutions, the reliance on non-contrast CT, and the absence of dynamic imaging assessment of hernia reducibility all constrain generalizability. The authors identify external validation on geographically diverse datasets, prospective comparison of automated measurements against intraoperative findings, and evaluation across hernia subtypes as the critical next steps before the system can influence operative decision-making. A recent scoping review identified only 20 relevant articles in AI-assisted hernia surgery, underscoring how much of this territory remains unmapped.

Subject of Research: Deep learning segmentation of ventral hernia CT scans for preoperative surgical planning

Article Title: Deep learning in ventral hernia imaging: automated multi-structure CT segmentation for surgical planning

Article References: Rengan, V., Meenashi Sundaram, P., Arora, E., Girieasen, S., Bawa, A., Alexander, N., Ravanasamudram Sitaraman, R., Reddy, V., Kalikar, V., Arora, A., Kona, L., Venkataramanan, R., Lalwani, D., Meenashi Sundaram, D., & Kalla, R. (2026). Deep learning in ventral hernia imaging: automated multi-structure CT segmentation for surgical planning. Journal of Abdominal Wall Surgery, 5, Article 15545. https://doi.org/10.3389/jaws.2026.15545

Image Credits: AI Generated

DOI: 10.3389/jaws.2026.15545

Keywords: deep learning, ventral hernia, CT segmentation, TransUNet, surgical planning, medical imaging, artificial intelligence, abdominal wall, hernia defect, IoU, proof-of-concept, component separation

The core problem the researchers set out to solve is one that anyone deploying artificial intelligence in the real world eventually confronts. Most high-performing image segmentation models depend on heavy backbone networks—deep stacks of convolutional layers borrowed from image classification research—combined with intricate feature-fusion schemes that merge information from many layers of the network. These designs excel at benchmark accuracy, but they impose punishing demands on memory, processing power, and energy. On an edge device with restricted computing resources, such a model may take seconds or minutes to produce a single prediction, drain a battery, or simply fail to fit in available memory. When the alternative is shipping raw images over a network to a cloud server for processing, new bottlenecks appear in the form of inference latency and communication cost, both of which matter enormously in time-sensitive applications such as clinical screening.

LMANet’s answer to this dilemma is an encoder-attention-decoder architecture that distributes the computational burden intelligently. At the front of the pipeline sits an efficient convolution-based compact encoder, whose job is to extract basic visual features from the input image. Because this encoder is deliberately kept small, it can run on the edge side—the device where the image is captured—without overwhelming local hardware. The subsequent stages of the network, a multiscale attention module that enriches the extracted features and a decoder that converts them into a final pixel-level segmentation, can then be executed flexibly by the cloud side whenever additional computational support is needed. This division of labor is the essence of cloud-edge collaboration: the edge does the lightweight work that must happen close to the data, while the cloud handles the heavier aggregation and decoding steps on demand.

The technical heart of the design is the multiscale attention module, and understanding why it matters requires a brief look at how segmentation networks see. An image contains structures at many different scales: a skin lesion, for example, has fine boundary details that must be traced precisely, a broader regional texture that distinguishes diseased tissue from healthy skin, and an overall semantic identity that tells the network what the object in question actually is. A network that processes everything at a single scale tends to miss some of these cues. Attention mechanisms address this by letting the model learn where to focus—weighting certain features more heavily than others based on their relevance to the task. LMANet’s multiscale attention module is designed to strengthen all three categories of information simultaneously: boundary details, regional contextual information, and global semantic responses, without requiring the enormous parameter counts that heavy fusion modules typically demand.

To evaluate the approach, the team turned to the ISIC 2018 dataset, a widely used public benchmark of dermoscopy images for skin-lesion analysis. Skin-lesion segmentation is a demanding test case for any lightweight model, because the clinical value of the output depends on the accuracy of the lesion boundary: a segmentation that slightly overestimates or underestimates the extent of a lesion can change a subsequent diagnosis or biopsy decision. The task therefore sits at the intersection of the two pressures the study aims to reconcile—pixel-level accuracy on one side and the practical constraints of inference latency, model size, and communication cost on the other. It is precisely the kind of workload that a cloud-edge visual analytics system would be expected to handle in a real deployment.

The reported results are striking for a model of this size. On the ISIC 2018 benchmark, LMANet achieved a Dice score of 0.921 and an intersection-over-union, or IoU, of 0.853. The Dice score, which ranges from zero to one, measures the overlap between the predicted segmentation and the ground truth, with values above 0.9 generally considered excellent in medical imaging. IoU, a stricter overlap metric that penalizes both false positives and false negatives, at 0.853 indicates that the model’s predicted lesion regions agree closely with expert annotations. Achieving these numbers is not unusual for large, computationally expensive networks—but LMANet does so with only 5.6 million parameters and 9.4 gigaflops of computation, figures that place the model firmly in the lightweight category suitable for resource-constrained hardware.

For context, many mainstream segmentation networks carry parameter counts several times larger, sometimes exceeding tens of millions of parameters and requiring substantially more floating-point operations per inference. Every additional parameter must be stored in memory, moved through the processor, and, in a cloud-edge setting, potentially synchronized between devices. Every additional floating-point operation translates directly into inference time and energy consumption. By holding the parameter count to 5.6 million while maintaining a Dice score above 0.92, LMANet demonstrates that the accuracy-versus-efficiency curve can be shifted favorably when the architecture is designed around the deployment environment rather than adapted to it after the fact.

The authors report that, compared with representative segmentation methods, LMANet shows a more balanced performance across the three axes that matter in practice: segmentation accuracy, model compactness, and inference efficiency. This notion of balance is arguably the study’s most important contribution. In the research literature, headline accuracy figures often dominate the conversation, but a model that wins a benchmark by a fraction of a percentage point while requiring ten times the compute may be a poor choice for an actual deployment. By explicitly optimizing for the joint objective, the researchers offer a template for how segmentation models for cloud-edge visual analytics might be evaluated going forward—not as isolated accuracy numbers but as profiles across accuracy, size, and speed.

The implications extend beyond dermatology. The architectural principles at work in LMANet—compact encoders at the edge, attention-driven multiscale feature aggregation, and flexible cloud-side decoding—apply to any visual analytics scenario where images are captured on constrained devices and analyzed with the help of remote infrastructure. Smart-city cameras, agricultural drones, industrial inspection systems, and remote diagnostic tools in underserved regions all face versions of the same constraint: the intelligence must be accurate, but it must also fit within tight budgets of latency, bandwidth, and energy. A model that can be split between edge and cloud, performing basic feature extraction locally and deferring heavier processing to the cloud only when needed, offers a practical pathway for such systems to adopt deep learning without rebuilding their hardware.

The study, published open access in the Journal of Big Data on 5 October 2026, was supported by the Doctoral Research Foundation of Weifang University, the Shandong Provincial Natural Science Foundation, and Wonkwang University. The authors, led by Yujie Li of Weifang University’s School of Computer Engineering with corresponding author Sun-Kyoung Kang of Wonkwang University’s Department of Computer Software Engineering, note that the results demonstrate the effectiveness of LMANet for computationally efficient skin-lesion segmentation and indicate its potential for resource-constrained cloud-edge visual analytics more broadly. As artificial intelligence continues to migrate from data centers into the devices that surround us, work like this suggests that the future of applied machine learning may belong not to the largest models, but to the cleverest ones—networks that know exactly where to spend their limited computation, and where to ask for help.

Subject of Research: Lightweight attention-based neural networks for efficient cloud-edge image segmentation of skin lesions

Article Title: Cloud-edge efficient image segmentation via lightweight multiscale attention learning

Article References: Li, Y., Zhang, H., Liu, G., Wang, T., Chen, C., & Kang, S.-K. (2026). Cloud-edge efficient image segmentation via lightweight multiscale attention learning. Journal of Big Data. https://doi.org/10.1186/s40537-026-01577-4

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01577-4

Keywords: image segmentation, cloud-edge computing, lightweight neural networks, attention mechanism, skin-lesion segmentation, ISIC 2018, Dice score, IoU, medical imaging, computer vision, deep learning, model efficiency