Efficient Deep Learning for Radiographic Body Part Classification

The growing demand for automated classification of medical X-ray images has driven the development of efficient deep learning models. This study introduces YOLOv8n-cls, a lightweight and accurate solution for body part classification in radiographic images, achieving a top-1 accuracy of 99.4%, outperforming several state-of-the-art models. A dataset of 22,500 images spanning 11 anatomical categories was constructed from public (MURA) and private DICOM sources, including previously underrepresented regions to improve generalizability. A comprehensive pre-processing pipeline comprising image conversion, windowing, and text removal via Keras OCR was employed, alongside various data augmentation techniques to enhance model robustness. Comparative evaluation with other deep learning models (EfficientNet B4, B2, B6, B0, DaViT Base, EfficientNetV2 XL) confirmed the superiority of YOLOv8n-cls in terms of both accuracy and computational efficiency. Model performance was assessed using top-1 accuracy, precision, recall, F1 score, loss curves, and confusion matrices. This work contributes a scalable, highperforming approach tailored for clinical deployment and highlights the importance of incorporating diverse anatomical coverage and pre-processing strategies for reliable X-ray classification.