DFU-Helper: An Innovative Framework for Longitudinal Diabetic Foot Ulcer Diseases Evaluation Using Deep Learning
Summary: Originally published in Applied Sciences (MDPI, 2023, 13(18):10310; DOI: 10.3390/app131810310) and recently archived on the HAL open science repository, DFU-Helper introduces a deep learning framework designed to address a practical gap in wound care: the absence of objective, longitudinal, computer-assisted assessment tools for diabetic foot ulcer (DFU) progression monitoring. With approximately 537 million people living with diabetes globally (projected to reach 783 million by 2045) and DFU representing a leading cause of lower limb amputation, continuous and meticulous patient monitoring is currently performed by medical practitioners on a daily basis — a resource-intensive process subject to inter-observer variability and lack of quantitative benchmarks between visits. DFU-Helper employs a Siamese Neural Network (SNN) architecture that learns feature-level similarity between DFU images across five distinct disease conditions: none, infection, ischemia, both (ischemia and infection combined), and healthy. At a patient’s initial consultation, an image is processed to compute distances from each class anchor point — generated using representative feature vectors — producing a comprehensive table and radar chart of disease-condition similarity distances. At subsequent visits, a new image is processed alongside the initial image, and DFU-Helper plots the progression trajectory, enabling visual and numerical comparison of disease state over time. Pseudo-labelling with a threshold of 0.9 yielded the best performance on the test dataset, achieving a Macro F1-score of 0.6455. The authors position DFU-Helper as a novel contribution distinguishable from prior DFU classification tools by its explicit focus on longitudinal follow-up rather than single-image diagnosis — to their knowledge, no existing tool at time of publication used deep learning comparably for DFU follow-up. The work was conducted collaboratively by researchers from Université des Mascareignes (Mauritius), XLIM/Université de Limoges (France), 3iL Ingénieurs (France), and Université de Limoges.
Key Highlights:
- Siamese Neural Network architecture trained on DFU image dataset; classifies five disease states: none, infection, ischemia, both (ischemia + infection), and healthy; best Macro F1-score 0.6455 using pseudo-labelling (threshold 0.9)
- Longitudinal assessment design: at initial visit, radar chart of class anchor distances generated; at subsequent visits, disease progression trajectory plotted — enabling objective numerical tracking between clinical consultations
- Clinical gap addressed: current DFU monitoring relies on daily practitioner visual assessment; DFU-Helper provides quantitative, reproducible, clinician-assistive output for longitudinal wound state tracking
- Five-condition classification covers the major wound state combinations relevant to DFU management — supports differentiated management decisions across infection, ischemia, combined, and clean wound states
- Pseudo-labelling technique: semi-supervised approach using high-confidence unlabelled samples (threshold 0.9) to expand effective training data — practically relevant given the limited scale of annotated DFU datasets
- Published in Applied Sciences (MDPI) 2023; deposited on HAL open science (hal-04554891v1) March 2026; open access CC BY 4.0; DOI: 10.3390/app131810310
Keywords: diabetic foot ulcer AI deep learning, DFU wound monitoring technology, wound care computer vision, diabetic foot infection ischemia classification, longitudinal wound assessment AI, neural network wound imaging
Mohammud Shaad Ally Toofanee, Sabeena Dowlut, Mohamed Hamroun, Karim Tamine, Anh Kiet Duong, Vincent Petit, Damien Sauveron