DEEP LEARNING IN WOUND CARE: SEGMENTATION OF WOUND AREA AND TISSUE IN DIABETIC FOOT ULCERS

Abstract

Diabetic foot ulcers (DFUs) are a serious complication for diabetes patients, often leading to lower limb amputation. Accurate segmentation of the wound area and its constituent tissues is crucial for effective treatment. This dissertation presents two novel segmentation approaches targeting different aspects of DFUs. The first approach focuses on wound area segmentation, introducing FUSegNet, an encoder-decoder architecture utilizing EfficientNet-b7 as its backbone. To overcome limited training samples, a modified spatial and channel squeeze-and-excitation (scSE) module, named parallel scSE (P-scSE), is proposed. Augmentations are applied for improved generalization. FUSegNet achieves a data-based Dice score of 92.70% on the Chronic Wound dataset and outperforms other scSE-based UNet models in Pratt's figure of merits (PFOM) scores on the FUSeg Challenge 2021 dataset, achieving a top-ranking dice score of 89.23%. The second approach targets tissue segmentation within DFUs, specifically focusing on fibrin, granulation, and callus tissues. With a limited dataset comprising only 110 labeled images and 600 unlabeled images, a semi-supervised learning (SSL)-based model is developed. A Mixed Transformer (MiT-b3) in the encoder and a CNN in the decoder are employed in the supervised phase, enhanced by a parallel spatial and channel squeeze-and-excitation (P-scSE) module. The semi-supervised phase employs a pseudo-labeling-based approach, iteratively incorporating valuable unlabeled images to enhance segmentation performance. The proposed method achieves a Dice score improvement from 84.89% in the supervised phase to 87.64% in the semi-supervised phase, outperforming state-of-the-art SSL approaches. These two approaches collectively advance the field of DFU segmentation, offering improved accuracy and efficiency in wound area and tissue segmentation, critical for effective treatment strategies.