Diabetic Foot Ulcer Grand Challenge 2021: Evaluation and Summary

TL;DR

This paper evaluates the 2021 Diabetic Foot Ulcer Challenge, highlighting dataset size, methods used, and performance results, with the top ensemble model achieving an F1-score of 0.6307.

Contribution

It provides a comprehensive evaluation and summary of methods and results from the Diabetic Foot Ulcer Challenge 2021, emphasizing dataset and model performance.

Findings

Largest publicly available diabetic foot ulcer dataset to date.

Top ensemble model achieved macro F1-score of 0.6307.

Highlights challenges and potential for semi-supervised learning.

Abstract

Diabetic foot ulcer classification systems use the presence of wound infection (bacteria present within the wound) and ischaemia (restricted blood supply) as vital clinical indicators for treatment and prediction of wound healing. Studies investigating the use of automated computerised methods of classifying infection and ischaemia within diabetic foot wounds are limited due to a paucity of publicly available datasets and severe data imbalance in those few that exist. The Diabetic Foot Ulcer Challenge 2021 provided participants with a more substantial dataset comprising a total of 15,683 diabetic foot ulcer patches, with 5,955 used for training, 5,734 used for testing and an additional 3,994 unlabelled patches to promote the development of semi-supervised and weakly-supervised deep learning techniques. This paper provides an evaluation of the methods used in the Diabetic Foot Ulcer Challenge 2021, and summarises the results obtained from each network. The best performing network was an ensemble of the results of the top 3 models, with a macro-average F1-score of 0.6307.