Deep learning in diabetic foot ulcers detection: A comprehensive evaluation.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic foot ulcers (DFUs) pose a significant health challenge.
• Automated detection using computer methods is crucial for early intervention.
• A systematic comparison of deep learning frameworks for DFU detection is lacking.

Purpose of the Study:

• To systematically compare state-of-the-art deep learning object detection frameworks for DFU detection.
• To evaluate the performance of various algorithms using the DFUC2020 dataset.
• To identify the most effective deep learning models for DFU recognition.

Main Methods:

• Comparison of deep learning algorithms: Faster R-CNN (and variants), YOLOv3, YOLOv5, EfficientDet, and Cascade Attention Network.
• Detailed description of model architectures, training parameters, pre-processing, data augmentation, and post-processing.
• Comprehensive evaluation using metrics such as mean average precision (mAP) and F1-Score on the DFUC2020 dataset.

Main Results:

• All evaluated methods required data augmentation and post-processing for optimal performance.
• Deformable Convolution, a Faster R-CNN variant, achieved the highest performance with an mAP of 0.6940 and an F1-Score of 0.7434.
• Ensemble methods improved F1-Score but did not enhance mAP.

Conclusions:

• Deep learning object detection frameworks show significant potential for automated DFU detection.
• Deformable Convolution is a highly effective approach for DFU recognition.
• Further research into ensemble methods could optimize F1-Score for DFU detection.