AUTOMATED SCALP DISORDER DIAGNOSIS USING ATTENTION-ENHANCED DEEP LEARNING WITH CLINICAL INTERPRETABILITY AND USABILITY VALIDATION

Abstract

Hair and scalp disorders affect millions of individuals worldwide, resulting in significant physical, psychological, and social consequences. Traditional diagnostic methods rely on subjective visual assessment by dermatologists, which is time-consuming, error-prone, and often inaccessible in underserved regions. This paper presents a machine learning framework for the automated diagnosis of scalp disorders using non-invasive dermoscopic and trichoscopic imaging. The proposed system employs a convolutional neural network (CNN) architecture based on fine-tuned ResNet-50, augmented with U-Net segmentation, Gaussian denoising, and CLAHE contrast enhancement. A dataset of 5,000 labeled scalp images was used for training and evaluation, with a 70/15/15 train-validation-test split. The system achieved an overall accuracy of 92.5%, precision of 91.3%, recall of 93.2%, F1-score of 92.2%, and AUC of 0.96 — surpassing prior published approaches. A user study with 10 dermatologists confirmed the system matched expert diagnostic accuracy in 95% of cases. This work demonstrates the transformative potential of deep learning-based dermatological tools for equitable, scalable, and accurate healthcare delivery.