A Review of Automatic Hair Removal in Dermoscopy Images: From Image Processing to Deep Learning

Comprehensive Summary

This article discusses the existing methods of removing hair artifacts from dermoscopy images. Bardou et al conduct a review and evaluation of conventional and novel hair removal methods and discuss their effectiveness in increasing diagnostic specificity. Conventional methods, which involve segmenting the hair and using the results to reconstruct the underlying lesion, showed a considerable percentage of false positives and residual hair pixels, receiving a Dice score of 53.00%. Novel methods, primarily deep learning methods, rely on generative models utilizing gate convoluted frameworks and datasets to detect and remove hair pixels. Deep learning methods showed to have higher performance in hair segmentation, with a Dice score of 96.88%. Deep learning methods therefore show considerably higher reliability than traditional methods.

Outcomes and Implications

The combination of dermoscopy with traditional investigation of skin lesions shows prospects of increasing diagnostic specificity and reducing false negatives or positives. However, hair occlusions in dermoscopy images has decreased diagnostic performance, as it can obstruct borders and texture. Shaving hair manually prior to imaging can damage skin tissue, creating an additional source of error. Researchers suggest using datasets (including SLICE-3D) and 3D total body photography to enhance reconstruction. Bardou et al emphasize the necessity of maintaining datasets to establish generalizability.