Deep Learning Based Cardiac MRI Segmentation: Do We Need Experts?

Deep learning methods are the de facto solutions to a multitude of medical image analysis tasks. Cardiac MRI segmentation is one such application, which, like many others, requires large number annotated data so that trained network can generalize well. Unfortunately, process having manually curated images by experts both slow and utterly expensive. In this paper, we set out explore whether expert knowledge strict requirement for creation sets on which machine successfully be trained. To do so, gauged performance three models, namely U-Net, Attention ENet, with different loss functions non-expert ground truth cardiac cine–MRI segmentation. Evaluation was done classic metrics (Dice index Hausdorff distance) as well clinical measurements, ventricular ejection fractions myocardial mass. The results reveal generalization performances neural is, all practical purposes, good data, particularly when receives decent level training, highlighting an opportunity efficient cost-effective annotations sets.