Automatic segmentation of regions of interest in hand radiographs using YOLOv10

Abstract

The objective of this paper is to present a methodology for the automatic segmentation of anatomical regions of interest in hand radiographs using the YOLOv10 object detection model. The segmentation was proposed as a preprocessing step for an automated bone age assessment system, allowing subsequent analyses to operate only on clinically relevant regions (joint, metacarpus, and carpus), thereby reducing the computational cost associated with processing full images. For model training, 204 images were manually annotated, divided into training and validation sets, and enriched using data augmentation techniques. The YOLOv10s model was evaluated using precision, recall, and overall accuracy metrics. The experimental results showed performance above 96% on the validation set and a global accuracy of 99.7% on the complete dataset of 18,787 images, with an average inference time of 0.11 seconds per image on CPU, demonstrating the practical feasibility of the proposed approach for clinical applications and computer-aided diagnostic systems.