Design and Numerical Analysis of Double Encoder-Swinnets-A Novel Swin Transformers-Based Diabetic Foot

Abstract

Thermography is crucial for early diabetic foot (DF) diagnosis and accurate segmentation of ulcer-prone areas. However, existing segmentation methods fall short due to image complexities and ambiguities. Recent advancements in deep learning show promise, but they rely on color images, not thermal ones. This research introduces an automated, robust, and precise diabetic foot segmentation approach using a deep neural network based on U-Nets and modified Swin transformers. The unique attention mechanism known as the axial attention parallel module (A2PM) is combined with the Unet-based Swin transformer model for an efficient segmentation process to extract local foreground features. The combination of the modified Swin Transformer’s multi-headed attention networks enhances thermal color information integration, resulting in superior segmentation accuracy. In addition, the proposed model makes use of the stacking dilated convolution (SDC) approach to protect the deep features that could be lost in the up-sampling modules. The feature maps are immediately integrated at the encoder and decoder stages using the shortcut connection (ResConv route) based on the residually connected convolutional layer. Furthermore, this ResConv path is added serially before the encoder and decoder characteristics are combined. The model is tested on 124 diabetic and 100 healthy subjects, evaluating its performance with metrics like DICE, IoU, precision, and recall. The suggested approach outperforms current techniques in an experimental evaluation, achieving 99.5% DICE, 98.9% IoU, 99.33% precision, and 99.56% recall for diabetic thermal ulcer image segmentation.