Hair and artifact removal in dermoscopic images using deep learning for enhanced skin cancer detection

The timely identification of melanoma and other cutaneous malignancies is fundamentally reliant on the acquisition of high-caliber dermoscopic images. However, factors such as hair interference, reflections, and various illumination artifacts frequently obscure essential characteristics of lesions, thereby impeding both clinical evaluations and automated diagnostic processes. This research introduces an innovative deep learning framework based on U-Net architecture, meticulously designed for the effective removal of hair and artifacts from dermoscopic images while ensuring the preservation of intricate lesion attributes, including pigmentary networks and irregular edges. The model underwent training predominantly on the HAM10000 dataset, which comprises 10,015 images sourced from multiple origins, and was rigorously validated on separate portions of both the HAM10000 dataset and the ISIC 2018 dataset. The training employed a hybrid loss function of Dice and Binary Cross-Entropy, alongside extensive data augmentation techniques and meticulous artifact mask creation. Quantitative assessments reveal significant improvements in image quality: Peak Signal-to-Noise Ratio (PSNR) increased from 21.5 dB to 34.1 dB, and Structural Similarity Index Measure (SSIM) enhanced from 0.79 to 0.89 for HAM10000 and from 0.77 to 0.92 for ISIC 2018, with Intersection over Union (IoU) ranging between 0.85 and 0.87 across the datasets. Subsequent melanoma classification utilizing a pre-trained model demonstrated notable improvements, with accuracy advancing from 84.2% to between 90.3% and 91.5%, F1-score rising from 81.6% to between 90.2% and 91.5%, along with an increase in prediction confidence. This methodology shows robust generalization capabilities across diverse artifact densities, types of lesions, and imaging environments, thereby establishing itself as a potent and adaptable preprocessing technique for standardized dermoscopic examinations and dependable AI-supported skin cancer diagnostics.