Dual-Stream VGG–DiNAT Framework With Red-Crowned Crane Optimization for Hair and Scalp Disorder Diagnosis

Abstract

In dermatological health care, accurate and reliable diagnosis of hair and scalp conditions is essential; however, traditional diagnostic processes often face challenges related to noise, variability in image quality and limited size datasets. To address these challenges, the present study describes a new Deep Learning (DL) framework that provides improved diagnostic accuracy. To offer effective noise suppression while retaining structural information, input images are augmented or enhanced by using a Spatially Adaptive Non-Local Means (SANLM) filtering based preprocessing model. An Edge-Guided Gaussian Mixture Model (EG-GMM) is then used for segmentation: the ability of the EG-GMM is to very accurately delineate the scalp and hair. To address limited data, a Progressive Growing GAN (PGGAN) model is used for creating realistic data augmentation to improve training variability. For robust classification, a dual stream classifier consisting of VGG19 and the Dilated Neighbourhood Attention Transformer (DiNAT) is used to maximise both long-range contextual attention and spatial feature extraction. For hyperparameter tuning of the framework RCCO (Red-Crowned Crane Optimisation) is employed to ensure improved classifier performance. The experimental validation in python software, demonstrates that the framework outperforms existing models in precision, accuracy, recall, FI-Score of 0.9929 and improved generalisation.