Scalp disease classification using modern and traditional deep learning architectures

Abstract

This thesis presents a benchmark study of seven deep learning architectures for classifying six scalp diseases: Healthy Scalp, Alopecia, Folliculitis, Dermatitis, Dandruff, and Hair Loss. Using a dataset of 1,368 images from Roboflow, the study evaluated CNN and Transformer models with and without data augmentation. ConvNeXt-B achieved 92.23% accuracy on the original dataset, while EfficientNetB2 achieved the highest accuracy of 93.69%, an AUROC of 0.9873, and an F1-score of 88.40% on the augmented dataset. The findings highlight the effectiveness of deep learning and data augmentation in improving scalp disease classification, particularly for small and imbalanced medical image datasets.