A Hybrid YOLO–CNN Model for Automatic Detection and Severity Assessment of Atopic Dermatitis in Infant Images

Abstract

Atopic dermatitis is one of the most common skin diseases affecting infants and children worldwide and has a particularly high prevalence in tropical countries. Traditional diagnosis methods, which still rely on physical examinations and laboratory tests often face challenges such as delays, high costs, and limited facilities, thereby necessitating an artificial intelligence–based system that is more efficient and accurate. This study aims to develop a hybrid YOLO–CNN model for the automatic detection and severity classification of atopic dermatitis in infants. The dataset comprises 2,000 infant skin images, including lesions categorized as mild, moderate, and severe, obtained from an online repository and field observations conducted in three villages. The labeling process was performed by a specialist doctor to ensure clinical validity. In the first step, YOLO was used to detect the lesion area in real time by generating a bounding box. This produced a region of interest (ROI), which was subsequently analyzed by a CNN model employing transfer learning in the second step to determine the severity level. Experimental results indicate that YOLO achieved high detection performance, with an mAP@0.5 of 91.2% and an F1-score of 90.2%, while the CNN model attained an average accuracy of 85% and a macro-F1 score of 85% in classification. The visualization of predictions indicates that most lesions were detected with confidence levels ≥0.9, confirming the model’s consistency. These findings highlight the potential of the hybrid YOLO–CNN framework as a supportive system for digital clinical diagnosis, applicable to both mobile applications and teledermatology services, particularly in regions with limited medical personnel. Future research should employ larger, multi-center datasets and integrate explainable AI approaches to promote broader clinical adoption.