Automated Cardiac Realtime MRI Evaluation

Abstract

We introduce our workflow to tackle automated evaluation of cardiac realtime MRI. The key approach is inspired by Active Learning and consists of N steps. First a limited amount of Training Data is annotated by staff with expert knowledge in the domain of pediatric cardiology. With this data we train a UNet using nnU-Net (Isensee, et. al). We then predict semantic labels with the trained model and use various techniques to judge the quality of each prediction. With that we are able to label each predicted segmentation with high or low quality. Predictions judged as low quality ones, are then presented to the domain experts and are manually corrected by them. Then, we can add those high quality labels to the training data set and start a new iteration by training the model. When the quality of predictions of an entire data set to be analyzed is high enough, we go on to synchronize the data set by assembling volumes of specific cardiac-respiration combinations based on the semantic segmentations. Finally, we are able to compute the stroke volume at different respiratory phases and compare them. The workflow explained above is deployed as a Plugin for the Software "3D Slicer".