Hair removal and lesion segmentation of dermoscopic images for classification of skin cancer using deep neural networks

Skin cancer is a prevalent health issue worldwide. Therefore, early detection through automated deep learning systems is crucial for saving lives. Hair presence in dermoscopic image presents diagnostic challenges by obscuring lesion features and complicating analysis due to variations in hair characteristics (density, color, and distribution), which can lead to diagnostic errors. In this work, a new and comprehensive approach is introduced to enhance the automatic classification of skin lesions. We propose an Efficient Hair Removal (EHR) technique that combines a Deep Residual U-Net with the TELEA inpainting algorithm, effectively eliminating hair artifacts from dermoscopic images. For precise lesion delineation, a Deep Residual U-Net model for skin lesion segmentation is also employed. The ISIC2019 dataset is used for skin lesion classification. Our approach progresses through five experimental stages, each building upon the previous. Starting with dataset balancing, which improved classification accuracy by 5%, we then applied our EHR framework, further boosting accuracy by 2.53%. The integration of skin lesion segmentation contributed to an additional 1.5% improvement. In the last, we use modified DenseNet169 architecture, which achieves a top accuracy of 97.74% on the ISIC2019 dataset, outperforming existing techniques. For lesion segmentation, Deep Residual U-Net achieved good results on the ISIC2018 dataset, with an Intersection over Union (IoU) of 0.8981 and a Dice Similarity Score (DSC) of 0.946.