A Deep Learning Approach to Horse Bone Segmentation from Digitally Reconstructed Radiographs

Abstract

Convolutional neural networks (CNN) are popular for segmentation and classification of bones in radiology. However, their training typically requires a database of thousands of manually segmented experimental images. In many cases, such a large dataset is not readily available in the community. In addition, manual segmentation is often too time intensive and prone to human perception, especially in cases of low image quality. In this paper, we show that a CNN can be accurately trained on the digitally reconstructed radiographs (DRR) of a 3D articulating shape model of the object of interest, bypassing the need for a manually-segmented database. The articulating model ensures a realistic appearance of the bones of interest in the DRR, thereby providing suitable training data for segmentation. As a proof-of-concept, we train a CNN on DRRs with the purpose of segmenting the phalanges of a horse leg from radiographs and show that it outperforms a geodesic active contour segmentation method in this particular case. Our proposed training procedure is effective for articulating objects and the resulting CNN can then be applied to real-data segmentation tasks, if preceded by appropriate augmentation.