A novel approach for hair removal in skin cancer images to enhance segmentation and classification performance.

This thesis focuses on the critical issue of skin cancer, highlighting the importance of early detection for better treatment and outcomes. Exposure to ultraviolet radiation is a significant risk factor, and there is a need for advanced detection methods. Using recent developments in computer vision, our study proposes a new preprocessing technique that specifically removes hair artifacts from skin cancer images. This technique aims to improve the accuracy of segmentation and classification. The method employs image processing techniques, including morphological operations and adaptive filters, to identify and isolate hair-containing regions. Subsequently, a reference mask guides a patch-based inpainting procedure, intelligently filling in hair-affected areas while preserving skin cancer lesion integrity. Evaluation on the International Skin Imaging Collaboration dataset involves comparing our approach with benchmark algorithms like "Dull Razor" and "E-Shaver." Deep learning classifiers, including Convolutional Neural Network based U-Net for segmentation and transfer learning classifiers (VGG16, InceptionV3, and EfficientNet), were trained using our hair-removed skin cancer images. Metrics such as accuracy, recall, precision, and F1-score assess our approach in segmentation, classification, and hair removal quality. Additionally, the quality of the hair segmentation is evaluated using Peak Signal-to-Noise Ratio, Intersection over Union, Mean Squared Error, Structural Similarity Index, and Universal Quality Index. Our proposed method significantly enhances both segmentation and classification, addressing challenges posed by hair artifacts in skin cancer images. These findings underscore the potential of our approach to advance skin cancer image analysis for clinical decision support systems.