FootSeg: Automatic Anatomical Segmentation of Foot Bones from Weight-Bearing Cone Beam CT Scans

This paper presented FootSeg, a novel method for automatic anatomical segmentation of the thirty-one foot bones from weight-bearing cone beam CT scans using deep neural networks. There were two challenges for the segmentation of foot bones in CT scans. The first was the scan variation, where the foot number and foot position differed in different scans. The second one was the severe class imbalance problem where the big bones, such as tibia and fibula, were much larger than the tiny bones like phalanges. We defined the FootSeg as a three-stage method, i.e., preprocessing, bone region segmentation, and bone pixel label classification, to solve the two challenges. The preprocessing step transferred the left/right-foot and two-feet scan into single right-foot scans to eliminate scan variation. The bone region segmentation module used a U-net model to extract the bone pixels, which acted as an indicator for the classification. In the bone pixel label classification part, we developed a patch-based CNN model to extract the local and global features of each bone pixel. We also designed a normalized spatial feature to represent the relative bone pixel position in the foot to further improve the classification model. The bone image feature and pixel position feature worked together to assign the correct bone label to each bone pixel. The proposed FootSeg method achieved a mean Intersection over Union of 90.3% over the thirty-one foot bones, demonstrating the method's outstanding performance. To the best of our knowledge, this was the first research of anatomical segmentation of all foot bones from weight-bearing CBCT using deep learning methods.