Synergistic Deep Learning Pipeline for Dermatological Image Classification: Cascaded U-Net Denoising and Feature-Selective CNN with Fused Feature Extraction

The utilization of skin disease images in diagnosis is increasingly necessitating high-quality visuals to ensure precise identification and treatment. Nevertheless, the presence of noise in these images poses a significant challenge by impeding diagnostic accuracy. To address these challenges, this work introduces a synergistic deep learning pipeline that combines denoising and feature refinement for improved skin disease classification. Initially, a cascaded U-Net architecture is utilized to eliminate noise from dermatological images, effectively preserving critical features such as lesion boundaries and textures. Following this, a feature-selective convolutional neural network (CNN) is applied to the denoised images, utilizing fused feature extraction to enhance the model's ability to identify disease-relevant patterns. This approach integrates Recursive Feature Elimination (RFE) with advanced feature fusion techniques to ensure optimal feature selection within the CNN framework. The proposed pipeline is evaluated using standard dermatological datasets and demonstrates improved classification accuracy compared to traditional methods. This dual-phase approach offers a robust solution for automating dermatological diagnoses while maintaining high diagnostic precision.