Unveiling and swift diagnosing chronic wound healing with artificial intelligence assistance

Abstract

Histopathological analysis of chronic wounds is crucial for clinicians to accurately assess wound healing progress and detect potential malignancy. However, traditional pathological tissue sections require specific staining procedures involving carcinogenic chemicals. This study proposes an interdisciplinary approach merging materials science, medicine, and artificial intelligence (AI) to develop a virtual staining technique and intelligent evaluation model based on deep learning for chronic wound tissue pathology. This innovation aims to enhance clinical diagnosis and treatment by offering personalized AI-driven therapeutic strategies. By establishing a mouse model of chronic wounds and using a series of hydrogel wound dressings, tissue pathology sections were periodically collected for manual staining and healing assessment. We focused on leveraging the pix2pix image translation framework within deep learning networks. Through CNN models implemented in Python using PyTorch, our study involves learning and feature extraction for region segmentation of pathological slides. Comparative analysis between virtual staining and manual staining results, along with healing diagnosis conclusions, aims to optimize AI models. Ultimately, this approach integrates new metrics such as image recognition, quantitative analysis, and digital diagnostics to formulate an intelligent wound assessment model, facilitating smart monitoring and personalized treatment of wounds. In blind evaluation by pathologists, minimal disparities were found between virtual and conventional histologically stained images of murine wound tissue. The evaluation used pathologists' average scores on real stained images as a benchmark. The scores for virtual stained images were 71.1 % for cellular features, 75.4 % for tissue structures, and 77.8 % for overall assessment. Metrics such as PSNR (20.265) and SSIM (0.634) demonstrated our algorithms' superior performance over existing networks. Eight pathological features such as epidermis, hair follicles, and granulation tissue can be accurately identified, and the images were found to be more faithful to the actual tissue feature distribution when compared to manually annotated data.