Deep Learning-Driven Heat Map Analysis for Evaluating thickness of Wounded Skin Layers

This paper introduces a non-invasive deep learning method for evaluating skin layer thickness in wounded sites, achieving 97.67% accuracy with ResNet18 and EfficientNet models. The technique uses a heatmap analysis to classify skin layers, including scars, wounds, and healthy skin, from a dataset of 200 annotated images. The study integrates CNNs and heatmaps for improved interpretability in medical AI, addressing limitations in previous approaches by enhancing accuracy, reducing overfitting, and providing a scalable tool for real-time monitoring. The ResNet18 model, with its residual learning mechanism, is best for model development, avoiding overfitting and improving generalization. The text also discusses an ablation study evaluating four pre-trained models, highlighting the critical role of learning rate in model training and the importance of fine-tuning hyperparameters for achieving optimal performance.