Using Anatomical Priors to Improve the Accuracy of the U-Net Deep Neural Network in Segmenting Medical Images of the Human Hand

The U-Net deep learning network architecture is generally effective in segmenting, or delineating, medical images into various regions, or classes, of interest. However, the segmentation performance of U-Net can be poor under certain circumstances, such as when different classes have sufficiently similar voxel properties and/or have overlapping spatial locations. While multiple plausible solutions to this problem are known, it is not possible to determine a priori which solution is appropriate for a given set of images, because the optimal solution depends greatly on image properties. In this brief report, we demonstrate that it is feasible to use the spatial probability maps, or 'anatomical priors', of various individual anatomical regions of the human hand to obtain highly accurate segmentation performance by the U-Net.