A Hybrid Distortion-Robust Skin Disease Classification Using Generative Preprocessing and Cross-Level CNN–Transformer Fusion

Abstract

Skin cancer is one of the most serious forms of cancer and melanoma is a very serious form if diagnosed too late. Although dermatoscopy is a non-invasive diagnostic tool for visualizing skin lesions, its diagnostic reliability is influenced by variability in image quality, overlapping visual features of the classes of disease diagnosis, and clinical interpretation subjectivity. In recent years, deep neural networks, particularly Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have shown promising results in assisting with the automatic detection and categorization of dermatological abnormalities. However, CNNs are limited in their ability to model global dependencies and ViTs' ability to generalize is limited by the amount of data. Furthermore, many existing methods are mainly evaluated against clean dermoscopic images, and do not explicitly consider robustness with respect to different image distortions.

To address these deficiencies, the current study introduces a hybrid framework combining generative preprocessing with CNN–Transformer cross-level fusion, which is robust to distortion. A generative adversarial network (GAN) is used to denoise, reduce blur and regularize illumination inconsistencies, while boosting the saliency of lesions. The CNN branch is used to extract local features and the ViT branch is employed to capture global relationship, then the two outputs are fused through cross-level fusion. The proposed approach outperforms the CNN-only, ViT-only and conventional hybrid approaches as shown by the evaluations performed on the HAM10000 dataset under clean and distorted conditions, and external cross-dataset validation on ISIC dataset. This work gives a major boost to the robustness of skin disease classification and provides a potential avenue for AI in dermatology.