An Intelligent Deep Learning Framework for Automated Alopecia Areata Detection and Severity Evaluation

Abstract

Objective: The goal of this research is to use scalp and hair images to create an efficient deep learning framework for the automated detection and severity categorization of Alopecia Areata (AA).
Methods: Using scalp and hair images collected from the Derm-Net and Figaro1K databases, the proposed ADAR-Net framework is developed to detect and assess Alopecia Areata. Important local and global characteristics are first extracted from the input images by two distinct FRCNN-based encoder networks. The model combines Cross-Residual Learning (CRL) and Multi-Task Learning (MTL) strategies to improve feature learning and enhance information transfer between related tasks. The relationships between the collected characteristics are then captured by combining them and passing them via an LSTM layer. However, the model is able to focus on the most important areas of the images while minimizing the impact of irrelevant data thanks to Channel Attention and Spatial Attention methods. The images are then classified into four categories: Healthy Alopecia Areata, Mild Alopecia Areata, Moderate Alopecia Areata, and Severe Alopecia Areata after the improved features are put into a fully connected layer and a SoftMax classifier.
Conclusion: The test findings demonstrate that, in comparison to all previous conventional systems, the proposed Attention-Driven Alopecia detection Network (ADAR-Net), which is designed to enhance the automated detection and classification of Alopecia Areata using hair and scalp image databases, achieves an accuracy of 95.1%.
Novelty: The combination of the CRL, MTL, LSTM, and attention processes in a single model for Alopecia Areata identification is what makes this work innovative. By using attention-guided learning, the suggested architecture concurrently learns complementing scalp and hair characteristics and improves their representation.