Can Deep Learning Wound Segmentation Algorithms Developed for a Dataset Be Effective for Another Dataset? A Specific Focus on Diabetic Foot Ulcers

Abstract

Diabetic foot ulcers (DFU) represent a severe complication, often resulting from poor glycemic control, neuropathy, peripheral vascular disease, or inadequate foot care. DFUs can lead to significant morbidity, including amputation and, in severe cases, can be fatal. Recently, advancements in computer vision technologies based on artificial intelligence (AI) have shown promise in DFU management. Particularly deep learning (DL) models such as U-Net and other models and techniques, were utilized to enhance wound segmentation accuracy. This research focuses on evaluating the generalization capabilities of DL models across different DFU datasets. Specifically, we investigated whether models trained on one dataset can be effective when utilized on another dataset, addressing the challenge of cross-dataset generalization. We employed 7 popular DL models, U-Net-VGG16, U-Net-EfficientNetV2S, ABANet, Ma-Net, LinkNet, DeepLabV3+, and Segment Anything Model (SAM), with 2 DFU datasets: FUSeg challenge and DFUC challenge. A total of 54 experiments were conducted plus 27 for SAM, involving training on one dataset, and testing on another, as well as training and testing on combined datasets. The results indicate substantial variability in segmentation performance when models trained on one dataset are tested on another, highlighting the influence of dataset characteristics on model generalization. The study underscores the importance of using diverse and comprehensive datasets to develop robust DL models for DFU segmentation and its generalization. This research contributes to the understanding of DL model performance in medical image segmentation and emphasizes the need for standardized datasets in improving DFU management through computer vision.

Keywords

• Wound segmentation
• deep learning
• diabetic foot ulcers