Automatic Scar Segmentation from DE-MRI Using 2D Dilated UNet with Rotation-Based Augmentation

Abstract

Delayed enhancement (DE) MRI is an important tool in diagnosis of cardiovascular disease as it can reveal different characteristics of the myocardium scars including infarction and no-reflow. Automatic segmentation of different regions has the advantage of improved accuracy and reduced inter-observer variability in quantifying key imaging biomarkers such as percentage of scars. In recent years deep learning has led to drastic performance improvement in automatic segmentation tasks using the UNet architecture. Cardiac MRI segmentation is a challenging task due to the high variability in imaging contrast, orientation and signal-to-noise ratio; specifically, for short-axis views, as they are double-oblique slices, due to the different orientations of the heart and operator choice, the overall appearance of the image can vary significantly, which poses a challenge to the neural networks as they cannot learn a consistent global anatomy. In this paper we developed a rotation-based augmentation to address this issue in both training and testing steps by eliminating the variance in orientations and demonstrated its effectiveness. 2D dilated UNet was used as the backbone network structure. On the test dataset, Dice scores of 0.933 for LV blood pool, 0.824 for myocardium, 0.578 for infarction and 0.697 for no re-flow regions were archived using the proposed framework.