Diabetic foot complications remain one of the leading causes of non-traumatic lower-limb amputation worldwide, yet current screening practices often rely on clinical examination that catches problems late. The prospect of a non-invasive, image-based AI tool that could flag at-risk feet before ulceration occurs represents a meaningful shift in preventive care for the estimated 537 million adults living with diabetes globally.
This systematic review, drawing from 2,214 candidate articles published between 2018 and 2023 and ultimately distilling nine high-quality studies, examined AI architectures applied to diabetic foot digital imaging. Thermal imagery — specifically plantar thermogram patterns reflecting abnormal temperature distribution — emerged as the dominant data modality. Deep learning approaches, chiefly artificial neural networks (ANNs) and convolutional neural networks (CNNs), were the most frequently deployed methods. The standout performer was an ANN model built on MATLAB's Image Processing Toolbox, achieving 97.5% classification accuracy across distinct macula types associated with diabetic foot pathology.
Placing this within the broader landscape, plantar thermography has been studied as a diabetic foot risk indicator for over two decades, with research showing that asymmetric foot temperature elevations can precede ulceration by days to weeks. What this review adds is a structured picture of how contemporary deep learning is now being systematically applied to that signal. However, the nine-study corpus is strikingly narrow, and the quality assessment using QUADAS criteria reveals the usual caveats of early-stage diagnostic AI: small, often single-center datasets, heterogeneous imaging protocols, and limited external validation. The headline 97.5% accuracy figure, while impressive, comes from a controlled experimental setting rather than a prospective clinical trial. For this technology to translate into standard diabetes care, larger multicenter validation studies and integration with electronic health record workflows will be essential. Overall, this is a promising but clearly incremental finding — a signal worth tracking rather than a practice-changing result.