A Fully Automated System for Localization and Classification of Foot Bones in X-Rays

Foot x-rays serve as a crucial diagnostic tool for identifying medical conditions affecting the foot. Among the critical steps in foot x-ray analysis are the localization and classification of individual bones. However, this task is often laborious and time-intensive. In this paper, we present a fully automated system for simultaneous localization and classification of foot bones. The system relies upon YOLOv7 convolutional neural networks-based object detection framework, and cascades two independently trained YOLOv7 networks: Backfoot-YOLOv7 and Forefoot-YOLOv7. Backfoot-YoLOv7 specializes in localizing and classifying larger bones in the hindfoot and midfoot (we termed backfoot), whereas, Forefoot-YOLOv7 specializes in localizing and classifying smaller bones in the forefoot. Experiments were conducted on a dataset comprising of 188 x-ray images. The results revealed that in comparison to applying YOLOv7 directly on the whole image (baseline approach), the proposed multi-step approach ensures overall improved performance. We achieved an overall accuracy of 0.98, with precision of 0.80, recall of 0.86, F1-score of 0.98, and IoU of 0.90. For the baseline approach we obtained the results of overall accuracy of 0.83, precision of 0.76, recall of 0.81, F1-score of 0.80, and IoU of 0.86.