Diabetic foot ulcers (DFUs) are a leading precursor to lower-extremity amputation, and outcomes depend heavily on how early an ulcer is recognized and how accurately its features, such as infection and ischemia, are judged. Because many people with neuropathy cannot feel early skin breakdown, and because specialist assessment is not always available, researchers have asked whether artificial intelligence (AI) applied to photographs of the foot could support earlier detection. This article reviews what published studies show about deep learning for DFU detection and classification, and where the evidence remains limited.
How image-based AI works in the diabetic foot
Most work in this area uses deep learning, a form of machine learning in which a neural network learns visual patterns from large sets of labelled images. In DFU research the tasks generally fall into three groups: detection (locating an ulcer within a photograph), classification (for example, labelling a wound as infected or ischemic), and segmentation (outlining the wound to measure its area). Systematic review evidence is early-stage. Tulloch and colleagues (IEEE Access, 2020) reviewed 37 studies applying machine learning to DFU prevention, diagnosis, and management, and found that studies commonly reported accuracy above 90% against reference standards. However, the authors concluded there was not yet enough evidence to confidently recommend these methods for routine healthcare use, calling for direct comparison with current standards of care, health economic analysis, and larger-scale data.
Lessons from the DFUC grand challenges
A major barrier has been the scarcity of large, shared datasets. The Diabetic Foot Ulcer Challenge 2020 (DFUC2020) addressed this by releasing 2,000 training and 2,000 test images and inviting teams to compete on ulcer detection. In the published comparison of leading methods, Yap and colleagues (Computers in Biology and Medicine, 2021) reported that the best-performing approach, a deformable-convolution variant of the Faster R-CNN architecture, reached a mean average precision of 0.694 and an F1-score of 0.743. All top methods depended on data augmentation and post-processing to reduce false positives, and ensembling several models improved F1-score but not mean average precision.
Classification is harder than detection. The follow-up DFUC2021 challenge asked teams to classify infection and ischemia in 15,683 wound image patches. Cassidy and colleagues (2022) reported a best macro-averaged F1-score of 0.631, achieved by an ensemble of the three top models. The authors noted that public datasets for this task are scarce and strongly imbalanced, meaning some wound types are far less represented than others. These results suggest that algorithms can reasonably localize an ulcer in a photograph but remain considerably less reliable at judging clinical features that determine treatment, such as infection and perfusion.
Early real-world testing
Cassidy and colleagues (Diabetes Research and Clinical Practice, 2023) reported a proof-of-concept clinical evaluation at two UK hospitals. A cloud-based model analysed 203 foot photographs, taken on low-cost Android smartphones, from 81 patients with diabetes; 162 images contained at least one ulcer. Using clinician ratings of the model’s predictions as the reference, sensitivity was about 0.92 and specificity about 0.89, rising to about 0.93 after a post-processing step that merged overlapping detections. Results were returned in roughly six seconds per image. Inter-rater agreement between clinicians was high (Krippendorff’s alpha 0.84 to 0.94).
The study also illustrated the technology’s limits. Participants were predominantly white, so performance across a range of skin tones is unproven. Testing occurred in hospital settings rather than in patients’ homes. The post-processing adjustment was applied retrospectively. The model also occasionally mistook other features, such as a partly visible toenail beside a wound, for an ulcer, reflecting limited diversity in the training data.
Current limitations
Across this literature, several recurring concerns temper enthusiasm. Datasets are modest in size and often come from a small number of centres. Class imbalance and variation in lighting, camera quality, and skin pigmentation can degrade performance outside the development setting. Many studies assess algorithms retrospectively on curated images rather than prospectively in routine care, and few have shown that using AI changes clinical outcomes such as healing time, hospitalization, or amputation. Reference standards are also variable, since clinician judgment of infection or ischemia from a photograph is itself imperfect.
Key takeaways
Deep learning can detect diabetic foot ulcers in photographs with promising sensitivity and specificity in early clinical testing, but classification of infection and ischemia remains less accurate. Current evidence comes mainly from technical challenges and small proof-of-concept studies. Larger, more diverse datasets and prospective trials measuring patient-centred outcomes are needed before image-based AI can be considered a validated adjunct to standard foot examination, which remains the established method for identifying the at-risk foot.
References
- Tulloch J, Zamani R, Akrami M. Machine learning in the prevention, diagnosis and management of diabetic foot ulcers: a systematic review. IEEE Access. 2020.
- Yap MH, Hachiuma R, Alavi A, et al. Deep learning in diabetic foot ulcers detection: a comprehensive evaluation. Computers in Biology and Medicine. 2021.
- Cassidy B, Kendrick C, Reeves ND, et al. Diabetic Foot Ulcer Grand Challenge 2021: evaluation and summary. In: Diabetic Foot Ulcers Grand Challenge 2021, Springer; 2022.
- Cassidy B, Yap MH, Pappachan JM, et al. Artificial intelligence for automated detection of diabetic foot ulcers: a real-world proof-of-concept clinical evaluation. Diabetes Research and Clinical Practice. 2023;205:110951.
Comments
