Deep learning based hair removal on ultraviolet-induced fluorescence dermatoscopy images

Ultraviolet-induced fluorescence dermatoscopy (UVFD) is a novel technique utilizing ultraviolet light to produce visible fluorescence. Many skin conditions reveal specific features during UVFD, guiding early detection and diagnosis. Although there are many methods for computer processing of dermatoscopic images, there are no methods targeting UVFD images. This work identifies newly encountered challenges and proposes a preprocessing pipeline for preparing obtained data for further statistical analysis.

The study discusses the challenges and the applicability of methods developed for regular dermatoscopic images. Finally, the universal pipeline for preprocessing is presented. The study used a collection of 2027 background images, manually selected by an expert from hairless regions, which were later used for hair generation. Additionally, 120 images were manually annotated for dark and light hair, resulting in a set of ground truth masks. The study evaluated a number of algorithmic methods and a deep neural networks: UNet, ChimeraNet and TransUNet.

The results clearly showed the superiority of deep learning over algorithmic methods in the task of hair segmentation in UVFD images. For dark hair segmentation, the UNet and TransUNet reported Dice scores ranging 0.95-0.98 on synthetic data. On real-life light hair UNet achieved the higher result, with Dice score of 0.557, while on dark hair all models ranged 0.61-0.65.

The study used a synthetic dataset to train the UNET neural network, which yielded results superior to analytical approaches known from dermatological images. The presented preprocessing pipeline is freely available and can be found on our GitHub repository https://github.com/aKempski01/UVFD-data-preprocessing.