Performance Analysis of Deep Learning Models for Hair Detection and Removal in Dermoscopic Images

Abstract

Removing hair from dermoscopic images is a key pre-processing step, because it directly affects the accuracy of later image analysis and diagnosis. This study implements, optimises, and evaluates three deep learning models for automated dermoscopic hair removal: a Multilayer Perceptron (MLP), a U-Net, and an Enhanced Attention Generative Adversarial Network (EA-GAN). Each model was tuned independently to its own best configuration, trained on a common dataset, and compared using six complementary image-quality metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Multiscale SSIM (MS-SSIM), and Visual Information Fidelity (VIF). U-Net performed best overall, with the highest PSNR, SSIM, MS-SSIM and VIF, which indicates superior structural preservation and perceptual fidelity. EA-GAN gave the lowest pixel-level reconstruction error but a comparatively low VIF, suggesting some loss of perceptual detail despite accurate pixel-level reconstruction. The fully connected MLP consistently trailed both encoder–decoder models, confirming its limited capacity to capture the complex spatial structure of hair artefacts. These findings indicate that convolutional encoder–decoder architectures with skip connections are best suited to dermoscopic hair removal and can meaningfully improve the reliability of fully automated skin-lesion analysis pipelines