Spatial-Frequency Feature Fusion: A Novel Deep Learning Architecture for Diabetic Wound Recovery

Abstract

The treatment of diabetic wounds presents a significant clinical obstacle. Traditional wound analysis methods are macroscopic and reliant on limited histological assessments. Biopsy sections stained with hematoxylin and eosin (H&E) and Masson’s trichrome are examined visually. This paper presents a plan for evaluating wound healing in diabetic patients. The classification model integrates a DiscreteWavelet Feature Extraction (DWTF) module for spatial-frequency feature extraction and a Spatial Feature Extraction (SFE) module; moreover, it shares information between the two modules to improve feature representation to ensure nuclear morphological features. Additionally, a segmentation method is developed to evaluate wound dimensions and contours for precise automated monitoring. To assess the efficacy of our method, we conducted histochemical and histopathological evaluations of full-thickness excisional wounds, utilizing wound biopsies fromWistar rats subjected to two distinct histochemical treatments.We utilized the identical deep neural network (DNN) architecture for both training and analysis, attaining a mean test set accuracy of 93.17%. We reassessed our evaluation of chronic wound healing by analyzing enhancements that improved our model’s performance. This resulted in a mean Dice score of 93.84% for segmentation and classification accuracy exceeding 90%. These findings underscore the promise of combining histopathological imaging with artificial intelligence.