Enhancing Wound Diagnosis with DoubleU-Net: An AI-Driven Segmentation Approach

Conventional wound assessment procedures rely heavily on the visual experience of skilled specialists to make diagnosis and treatment recommendations. However, a lack of competence can result in an incorrect diagnosis of wound etiology, ineffective wound care, and inadequate record keeping. Automating the segmentation of wounds in natural images plays a crucial role in the diagnostic and treatment process, enabling precise wound evaluation and providing essential quantitative data for therapy. By integrating an advanced segmentation algorithm into clinical workflows, it becomes possible to measure wound regions accurately and extract valuable insights for better treatment planning. Deep learning techniques have demonstrated remarkable success in medical image processing, particularly in semantic segmentation. This study employs a dual-stacked U-Net architecture (DoubleU-Net) with ASPP and squeeze-and-excitation blocks for diabetic foot ulcer (DFU) segmentation. The first U-Net incorporates a pretrained VGG-19 encoder, utilizing ImageNet features for efficient transfer learning. Then, a second U-Net is added to effectively gather more semantic information, and an Atrous Spatial Pyramid Pooling (ASPP) layer is used to increase contextual feature extraction throughout the network. The model was trained in the diabetic Foot Ulcer Challenge Dataset, which is publically available, and its performance was tested using the AZH wound database. The suggested approach outperformed state-of-the-art segmentation networks with a Dice score of 94.20%, as demonstrated by experimental findings.