Fully automated segmentation of foot bones using machine learning and convolutional neural networks.

Background: Recent advancements in medical imaging technology have significantly increased the prevalence of automatic and semi-automatic segmentation techniques for foot bones, offering promising potential for improving diagnostic accuracy and efficiency. However, a critical challenge remains the scarcity of literature on the reliability and validation of these automated systems, underscoring the need for comprehensive studies to ensure their trustworthiness in clinical practice.
Purpose: To implement a fully automated foot bone segmentation method processed exclusively using convolutional neuronal networks (CNNs).
Material and Methods: Foot bones of 50 computed tomography (CT) scans were manually segmented. Of them, 48 were used to train three CNNs of a customized and optimized three-dimensional (3D) U-Net structure for the segmentation process. The so trained networks were then applied on the remaining two CT scans. The Dice coefficient and the Intersection over Union (IoU) metric were calculated to evaluate the CNN's ability of proper foot bone segmentation.
Results: The CNN accurately segmented 5,090,689/5,434,749 voxels in the test sets, achieving an overall Dice coefficient of 0.97 and IoU of 0.94. Excellent segmentation results were obtained for the hindfoot, midfoot, hallux, sesamoids, and proximal phalanges, while lower performance was noted for the intermediate and distal phalanges of the lesser toes.
Conclusion: The CNN networks demonstrated excellent ability to recognize foot bone structures on CT. Our findings underscore the potential of deep learning models in providing reliable and accurate segmentation of foot bones, paving the way for more widespread clinical adoption.