Automatic quality control in computed tomography volumes segmentation using a small set of XCAT as reference images
Deep learning methods have performed superiorly to segment organs of interest from Computed Tomography images than traditional methods. However, the trained models do not generalize well at the inference phase, and manual validation and correction are not feasible for large-scale studies. Therefore, automatic methods to detect segmentation failure are crucial for Computer Aided Diagnosis systems. In this work, we present an automatic quality control method that can be used to reject poor segmentation. We register new test cases against a set of XCATreference or training images. This "reverse classification accuracy"approach uses similarity of image registration to estimate segmentation quality. We validated this approach on two large public CT datasets, CT-ORG and ABDOMEN-1K with multiple organs. We show empirical cutoffs for predicted similarity coefficient for organs of interest in public datasets that can be used for datasets where ground truth is not available.