A computer-based framework leveraging deep learning was developed for automated skin disease diagnosis, addressing the inaccuracies and inconsistencies of traditional manual methods. The system employs a two-stage process. First, an Adaptive Refined UNetV4 (ARUNetV4) performs disease segmentation by focusing on fine-grained lesion details while suppressing noise. The ARUNetV4's hyperparameters are optimised using the enhanced random variable-based red panda optimisation (ERV-RPO) algorithm. In the second stage, the segmented images are classified using a hybrid Vision Transformer with Residual DenseNet (ViT-RDNet). This model combines ViT's global contextual understanding with RDNet's local feature extraction to overcome visual similarities between different diseases. The framework demonstrated superior performance against existing models, achieving 96% accuracy on Dataset-1 for classification and 95.04% accuracy on Dataset-2 for segmentation.
Jayakanth et al. (Thu,) studied this question.