DEEPDERMANET: A Novel Dermatology Lesion Diagnosis Model Using Deep Learning Techniques

The increasing prevalence of skin diseases, driven by factors such as UV exposure, pollution, and lifestyle modifications, underscores the importance of proper skin lesion classification in dermatology. Though Deep Learning (DL) models are adept at identifying intricate patterns from dermoscopic images, they tend to have issues like high computational complexity, overfitting, poor interpretability, and large amounts of annotated data requirements. To deal with these problems, the current study suggests DeepDermaNet, an overall framework for accurate skin lesion diagnosis. The methodology starts with image pre-processing with DullRazor hair removal, CLAHE to enhance the contrast, and the Gray World algorithm to normalize colors. Segmentation of lesions is carried out with the suggested DermaSegNet-RCNN to segment Regions of Interest (ROIs). Feature extraction involves color histograms, color moments, elliptical Fourier descriptors (EFDs), and deep features with DermaEffiNet. To avoid overfitting, Recursive Feature Elimination (RFE) is used. The final DeepDermaNet model combining ResNet, attention mechanism, and convolutional layers has a high accuracy of 98.98%, outperforming current methods and holding clinical potential.