Led by Hadil Aldhubiea of Qatar University, together with colleagues in Qatar and Bangladesh, the review followed the PRISMA 2020 reporting standard and combed through PubMed, IEEE Xplore and Scopus for studies published between January 2019 and December 2024. From 141 records screened at the title-and-abstract stage, the team ultimately included 48 distinct deep learning study configurations drawn from 46 publications, spanning the full spectrum of diabetic foot ulcer analysis: image classification, wound grading, lesion detection, pixel-level segmentation, and multimodal fusion of different data streams. The search covered everything from convolutional neural networks and EfficientNet backbones to transformer hybrids, Siamese networks, generative adversarial networks and recurrent architectures applied to longitudinal patient data.

The technical landscape the reviewers mapped is dominated by convolutional neural networks, which appeared in roughly 60 percent of the included models. These networks learn hierarchical visual features, from edges to textures to wound patterns, and are well suited to capturing the fine-grained cues that matter in ulcer imagery, such as irregular wound edges, granulation tissue and the redness of surrounding skin. Transfer learning from large natural-image datasets like ImageNet remains the workhorse strategy, because most diabetic foot ulcer datasets are simply too small to train deep networks from scratch. More recent designs push further: DFU-SIAM combines an EfficientNetV2 backbone for local feature extraction with a BEiT transformer for global context, paired with a Large Margin Cotangent Loss to separate clinically similar categories under class imbalance, while DFU VIRNet fuses visible and infrared streams to catch physiological signals that ordinary photographs miss.

That multimodal angle is one of the most promising threads in the literature. Thermal imaging can reveal temperature gradients across the plantar surface that reflect inflammation, ischemia and neuropathic risk, potentially flagging danger before any wound is visible to the naked eye. The review found that within individual studies, multimodal pipelines combining RGB with thermal data beat their own unimodal baselines by a median of about 2.6 percentage points in classification accuracy and 3.1 points in Dice score for segmentation. Those gains are real but modest, and the authors caution that they are correlational, confounded with dataset choice, acquisition hardware and validation protocol. More troubling, many studies omit the engineering details, geometric registration, radiometric calibration, synchronization and normalization, that determine whether a fusion pipeline can be reproduced at another site or on another camera.

To move beyond anecdote, the reviewers built a 13-criterion quality rubric and computed a Weighted Methodological Rigor Score for every study, giving the heaviest weight to external validation, the single strongest indicator that reported performance is not an artifact of one dataset. A sensitivity analysis across five alternative weighting schemes confirmed the rankings were robust, with Spearman correlations of at least 0.92. The results were unflattering: the mean weighted score was just 49.0 percent, ranging from 23.8 to 95.2 percent. Only 27.1 percent of studies reported any external validation, a mere 8.3 percent released code, and just 16.7 percent discussed mobile or web deployment constraints. Just over half used publicly available datasets, with the Kaggle DFU collection and the DFUC grand challenge series the most common.

The most striking finding is a simple cross-analysis with big implications. Among the 36 configurations reporting a headline classification accuracy, the 11 studies that had undergone external validation reported a mean accuracy of 91.0 percent, while the 25 studies evaluated only on internal splits reported 94.3 percent. That gap of roughly 3.3 percentage points is exactly what you would expect if internal-only evaluation inflates performance, and it aligns with a broader pattern: 31 of those 36 configurations cluster between 90 and 99 percent accuracy, a suspiciously tight and optimistic distribution consistent with positive-outcome publication bias. The authors are careful to frame this as descriptive rather than causal, but the message is hard to ignore, headline accuracies in this field likely overstate what models will do in the real world.

Underlying many of these problems is the data itself. The review documents how lesion-to-image area ratios vary by more than two orders of magnitude across datasets, meaning a model tested on tightly cropped 150-by-150-pixel patches where the ulcer fills the frame faces a fundamentally easier task than one analyzing full-foot photographs where the wound occupies less than 5 percent of the image. Class imbalance was explicitly reported in 37.5 percent of studies, regional bias in 45.8 percent, and limited demographic diversity in 16.7 percent. Patch-based pipelines introduce a further hazard: if patches from the same patient appear in both training and test sets, leakage silently inflates results. Only about 62.5 percent of studies reported complete demographic information, and no publicly available dataset currently distributes paired images and clinical metadata under a redistributable license, a gap the reviewers identify as a priority for patient-aware modeling.

Evaluation practice compounds the confusion. Accuracy alone is misleading when severe categories like ischemia and gangrene are rare, and half of the included studies framed the task as binary ulcer-versus-healthy classification, which the authors argue has limited standalone clinical value because treatment decisions hinge on infection, ischemia, depth and Wagner grade. Segmentation studies report Dice and intersection-over-union inconsistently, detection papers often omit the IoU thresholds underlying their mean average precision figures, and pixel accuracy can look impressive simply because ulcers occupy so little of the image. The review proposes a minimum reporting checklist, including confusion matrices, macro-averaged metrics, boundary-aware segmentation measures, explicit patient-level splits and confidence intervals, as a baseline for future work.

The deployment picture is equally underdeveloped. Real-world use, especially home monitoring by patients with reduced foot sensation and possible visual impairment, demands pose-tolerant detection, robustness to mixed indoor lighting, and in-app capture guidance. Thermal pipelines carry their own discipline: plantar temperature is sensitive to recent footwear, ambient conditions and activity, and established protocols require at least 15 minutes of foot acclimatization in a controlled room at roughly 22 to 24 degrees Celsius before imaging. Models trained on acclimatized data will drift out of distribution the moment those conditions are not met, which is why the reviewers recommend contralateral-foot reference checks and restraint about high-confidence predictions on out-of-protocol captures. Not a single one of the 48 studies provided a complete specification for integration with electronic health records, a translational gap the authors call substantial.

The review’s closing roadmap is clear-eyed rather than defeatist. It calls for leakage-resistant benchmarking with patient-level separation and cross-dataset testing, standardized acquisition and calibration guidance for thermal imaging, lightweight and energy-aware model design for edge devices, clinically interpretable explainability extending to modality-specific contributions, and open release of code, weights and evaluation pipelines. The authors’ central conclusion reframes the field’s problem: progress in diabetic foot ulcer AI is currently limited less by the availability of powerful architectures than by the absence of standardized, transparent and clinically grounded evaluation. Until external validation, honest reporting and reproducible engineering become routine, the impressive accuracy figures filling the literature should be read as upper bounds, not guarantees, of what these systems will deliver at the bedside.

Subject of Research: Deep learning methods for diabetic foot ulcer detection, segmentation and multimodal assessment

Article Title: Deep learning for diabetic foot ulcer: a systematic review of segmentation, classification and multimodal integration

Article References: Aldhubiea, H., Fetais, N., Newaz, M., Islam, M. S. B., Chowdhury, M. E. H., Suganthan, P. N., & Kiranyaz, S. (2026). Deep learning for diabetic foot ulcer: a systematic review of segmentation, classification and multimodal integration. Cluster Computing, 29(14), Article 826. https://doi.org/10.1007/s10586-026-06619-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06619-3

Keywords: diabetic foot ulcer, deep learning, systematic review, convolutional neural networks, image segmentation, thermal imaging, multimodal fusion, external validation, reproducibility, medical imaging, clinical deployment, PRISMA

The research team, led by Yman Chemlal of Hassan II University in Casablanca together with colleagues spanning computer science laboratories and the neurology department of the Mohammed VI University Hospital Center in Tangier, Morocco, applied PRISMA guidelines to select and analyse a decade of literature. Rather than simply cataloguing complaints about AI in epilepsy, as earlier reviews had done, the team took a quantitative approach. They examined how research is actually conducted across four clinical applications: seizure detection, seizure prediction, diagnosis, and treatment support. For each study they recorded the datasets used for training and testing, the data modalities involved, the model architectures chosen, and the evaluation metrics reported. They then compared these observations with the obstacles most frequently cited in the literature, searching for the true root causes of what researchers call the translational gap.

The diagnosis that emerged is stark. The authors identified five interconnected issues that together explain why AI in epilepsy remains largely confined to research papers. The first is a heavy reliance on public datasets, which the review found fail to capture the complexity, noise, and multimodal character of real-world clinical data. Publicly available electroencephalography collections are typically curated, relatively clean, and recorded under controlled conditions. Genuine clinical EEG, by contrast, is riddled with muscle artefacts, electrode faults, patient movement, and the sheer variability of dozens of different recording devices and protocols across hospitals. A model that achieves near-perfect accuracy on a benchmark dataset can collapse when confronted with the messy reality of a hospital monitoring unit, a phenomenon well documented in the broader machine learning for healthcare literature.

The second finding is a striking imbalance in what the field chooses to work on. The review found that 84 percent of studies focus on seizure detection and prediction, and these are precisely the applications dominated by public datasets that were primarily designed for identifying seizure events in the first place. The consequence is a self-reinforcing loop: the easiest data to obtain shapes the research questions that get asked, which in turn produces yet more papers on the same narrow set of tasks. Meanwhile, applications with arguably greater clinical value, such as supporting treatment decisions or assisting diagnosis of difficult cases, remain comparatively underexplored, starved of the large, well-annotated datasets that would make them tractable.

Third, the review uncovered a fundamental mismatch in how success is measured. Most research evaluates model performance using technical classification metrics such as accuracy and sensitivity, which together account for 85 percent of all metrics used across the analysed studies. These numbers are meaningful to machine learning engineers but say little about whether a model actually helps a clinician or a patient. A seizure detection algorithm with 99 percent accuracy might still generate so many false alarms that it becomes unusable in a long-term monitoring setting, or it might miss the seizures that matter most clinically. The authors argue that the field needs evaluation frameworks anchored in genuine clinical needs, such as the number of false alarms per day for wearable devices or the effect on treatment outcomes, rather than abstract percentages computed on held-out test data.

The fourth issue concerns the very architectures that have driven AI’s recent success. The emergence of powerful and high-performing models, including convolutional neural networks and elaborate hybrid designs that combine convolutional, recurrent, attention, and graph-based components, has brought impressive benchmark results but also serious interoperability problems and the notorious black box problem. When a deep network flags a segment of EEG as epileptic, it rarely explains why in terms a neurologist can verify. The review’s nuanced conclusion here is that the challenge is not interpretability in the abstract but clinician reluctance rooted in a lack of trust in these models. Explainable AI techniques have been proposed for seizure detection on EEG signals, but unless the explanations align with clinical reasoning and are validated with the clinicians who must act on them, trust will not follow.

Perhaps most importantly, the fifth finding reframes the translational gap not as a single wall but as a set of different barriers depending on where you stand. In seizure detection and prediction, the dominant obstacles are algorithmic: models that do not generalise beyond their training data and metrics that do not reflect deployment conditions. In diagnosis, the bottleneck is data: too few large, diverse, well-labelled datasets to train and validate models that can distinguish epilepsy from its many mimics. In treatment support, the challenge shifts again to deployment: even promising tools for predicting which antiseizure medication will work for a given patient face regulatory, integration, and workflow hurdles that computer science alone cannot solve. Any serious roadmap, the authors contend, must therefore be tailored application by application rather than prescribing one-size-fits-all fixes.

Based on this multidimensional diagnosis, the team proposes a specific roadmap with strategic actions for each clinical application, aiming to align AI development with actual clinical needs. The directions they sketch echo recommendations already gaining traction across medical AI more broadly, including reporting standards such as TRIPOD+AI for clinical prediction models, the CLAIM checklist for AI in medical imaging, and the DECIDE-AI guidelines designed specifically to bridge the development-to-implementation gap. The review also provides what the authors describe as a scientific framework for guiding the next decade of AI research in epilepsy, intended to help researchers, funders, and clinicians prioritise work that is most likely to reach patients.

The stakes are considerable. Epilepsy affects tens of millions of people worldwide, and roughly a third continue to have seizures despite medication. For these patients, reliable seizure forecasting could transform daily life, restoring the ability to drive, work, and live independently. AI-assisted diagnosis could shorten the often years-long delay before a correct diagnosis and appropriate treatment, particularly where access to specialist neurologists is limited. Tools that predict medication response could spare patients years of trial and error with drugs that fail. None of these benefits will materialise, however, if the research community continues to optimise models for benchmarks rather than for clinics.

The review, published on 23 September 2026 as volume 38, article 751 of Neural Computing and Applications, is likely to become a reference point for the field precisely because it replaces anecdote with measurement. By quantifying where the decade’s effort has gone, which datasets and architectures dominate, and which metrics are used to declare success, it gives the epilepsy AI community its clearest picture yet of why the promised revolution has stalled. The message to researchers is uncomfortable but constructive: the bottleneck is no longer model performance on curated data. It is the alignment between what algorithms are trained to do and what patients and clinicians actually need. Closing that gap, the authors argue, will require clinicians and computer scientists to design studies together from the outset, evaluate models against clinical endpoints, and confront the trust problem head on. The next decade of AI in epilepsy will be judged not by leaderboard scores but by whether algorithms finally begin changing outcomes in the clinic.

Subject of Research: Barriers to translating artificial intelligence applications from research into clinical practice in epilepsy

Article Title: The translational gap in AI for epilepsy: a systematic review and recommendations for future research

Article References: Chemlal, Y., Benamri, I., Ferjouchia, H., Ferjouchia, Z., & Rabhi, A. (2026). The translational gap in AI for epilepsy: a systematic review and recommendations for future research. Neural Computing and Applications, 38(18), Article 751. https://doi.org/10.1007/s00521-026-12460-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12460-x

Keywords: artificial intelligence, epilepsy, translational gap, systematic review, seizure detection, seizure prediction, electroencephalography, deep learning, clinical deployment, machine learning, neurology, explainable AI

The historical roots of this transformation run deeper than many patients realize. As early as 1996, researchers were using artificial neural networks to score radiotherapy treatment plans, and by the turn of the millennium similar networks were being used to customize beam orientations and to predict biological outcomes in prostate cancer. What has changed is the scale and sophistication of the tools. Deep learning, a subfield of machine learning in which neural networks learn directly from raw data, has enabled convolutional neural networks to segment organs at risk in head and neck CT images with a fidelity that now rivals, and in some cases exceeds, manual delineation by experienced clinicians.

Auto-contouring is perhaps the most mature application. Commercially available and, in many jurisdictions, regulator-approved tools now outline organs such as the heart, spinal cord and parotid glands automatically, and multi-center evaluations of deep learning CT autosegmentation in the head and neck region have demonstrated clinically acceptable performance. A randomized controlled crossover trial evaluating AI-assisted heart contouring, along with prospective multi-center studies of AI-assisted delineation in postoperative lung cancer radiotherapy, reflect a growing body of clinical evidence rather than mere laboratory promise. Yet the review is careful to note that expert human oversight remains essential across AI-enabled workflows, and that the tendency of users to over-rely on automated recommendations, a phenomenon known as automation bias, is itself a documented risk that clinics must actively manage.

Treatment planning, historically the most labor-intensive stage of the radiotherapy workflow, is being transformed along parallel tracks. Knowledge-based planning systems predict dose-volume histograms for organs at risk and have been extended to fully automated planning across ten different cancer sites. Deep reinforcement learning, in which models learn to make decisions by optimizing actions based on feedback from an environment, has produced proof-of-principle systems for high-dose-rate brachytherapy in cervical cancer and hierarchical frameworks for prostate and head and neck intensity-modulated radiation therapy. In parallel, generative AI approaches are being applied to synthetic computed tomography generation, allowing dose calculations to proceed from MRI scans alone, an advance with particular relevance for MR-guided adaptive radiotherapy, where dosimetric evaluations using cycle-consistent generative adversarial network synthetic CT have shown feasibility.

Adaptive radiotherapy, in which plans are modified during a treatment course to account for anatomical change, represents another arena in which AI is already operational. Magnetic resonance-guided prostate stereotactic body radiation therapy with daily online plan adaptation has been tested in prospective phase 1 trials, and multi-institutional phase 2 trials of ablative five-fraction stereotactic MR-guided on-table adaptive radiotherapy for pancreatic cancer have demonstrated the clinical viability of rapid, AI-assisted replanning. Because breathing and other motions can shift targets by significant margins, the ability to re-contour and re-plan in minutes rather than hours changes what is therapeutically achievable, particularly for tumors near the gastrointestinal tract.

Beyond the machinery of planning and delivery, the review highlights AI’s expanding role in predicting outcomes. Models trained on imaging, dose and clinical data can forecast toxicities such as radiation pneumonitis, xerostomia and bowel toxicity, with recent radiomics and dosiomics machine learning approaches drawing on the prospective multicenter RTOG 0617 and REQUITE trials to predict symptomatic lung injury. Multimodal AI, which integrates imaging, text, genomics and clinical data into unified predictions, is extending this capability further, and transformer-based modeling of patient-reported outcomes is being explored for survival prediction in non-small cell lung cancer. Large language models are even being benchmarked for extracting toxicity data from oncology trial records, while ambient AI scribes are being tested to relieve the documentation burden that contributes to clinician burnout.

The most forward-looking section of the review concerns the convergence of digital twins, foundation models and agentic AI. A digital twin is a computational replica of an individual patient that can be used to simulate treatment scenarios before they are applied, and predictive digital twins have already been explored for optimizing patient-specific radiotherapy regimens under uncertainty in high-grade gliomas, as well as for adaptive proton therapy strategies using cone-beam CT. Foundation models, large-scale models trained on broad and diverse datasets that can be adapted to downstream tasks with minimal task-specific training, could serve as the engine of such twins, while agentic AI systems composed of multiple autonomously orchestrated agents could coordinate the complex interdependent steps of a full radiotherapy workflow. Feasibility studies of automating radiotherapy planning with large language model agents suggest this is no longer speculative.

The authors are equally candid about why deployment has stalled. Data leakage, in which information from validation or test data inadvertently informs model training, produces misleadingly optimistic performance estimates. Silent model failure, in which performance deteriorates without warning as patient populations or clinical practices shift, poses a patient-safety hazard that conventional quality assurance was never designed to catch. Dataset biases can propagate into population-level disparities in autocontouring accuracy, and hallucination, the generation of fabricated outputs ungrounded in reality, remains an inherent hazard of generative systems. Explainable AI and uncertainty estimation methods that flag low-confidence predictions for human review are proposed as partial remedies, alongside failure mode and effects analyses of automated tools and quality assurance frameworks specifically designed for AI-based applications, including deep learning approaches that predict gamma passing rates for patient-specific QA.

Regulatory and governance infrastructure is beginning to catch up. A joint ESTRO and AAPM guideline for the development, clinical validation and reporting of AI models in radiation therapy has been published, the US Food and Drug Administration has issued guidance on predetermined change control plans for AI-enabled device software and good machine learning practice principles, and the European Union’s AI Act establishes a broader legal framework. Federated learning, which trains models across institutions without moving patient data, is emerging as a technical answer to privacy constraints, with demonstrated applications in dose-volume parameter prediction and survival modeling for hepatocellular carcinoma. Professional societies are also pressing for structured education, including national workshops on training the next generation of radiation oncologists in AI and centralized databases of AI courses in Europe.

What emerges from the review is a field at an inflection point. Ongoing clinical trials are now evaluating not just the technical performance of AI-enabled applications but their safety, efficacy and genuine clinical value, a shift from proof-of-concept toward accountability. The authors, whose work is supported by the National Institutes of Health and the Department of Defense, argue that the convergence of digital twins, generative AI and foundation models is poised to shape the next era of AI-enabled radiotherapy, one in which treatment could be continuously simulated, personalized and adapted for each patient. But they temper that optimism with a clear-eyed warning: without practical recommendations for responsible adoption, rigorous validation and sustained human oversight, the technologies capable of optimizing the delivery of radiotherapy will remain, in clinical terms, underused. The challenge, in other words, is no longer building the intelligence; it is building the trust.

Subject of Research: Artificial intelligence applications in radiation oncology treatment planning, delivery and outcome prediction

Article Title: Optimizing the delivery of radiotherapy with artificial intelligence

Article References: Katsoulakis, E., & El Naqa, I. (2026). Optimizing the delivery of radiotherapy with artificial intelligence. Nature Reviews Clinical Oncology. https://doi.org/10.1038/s41571-026-01204-4

Image Credits: AI Generated

DOI: 10.1038/s41571-026-01204-4

Keywords: artificial intelligence, radiotherapy, machine learning, deep learning, digital twins, generative AI, auto-contouring, treatment planning, adaptive radiotherapy, foundation models, outcome prediction, clinical deployment