Using MRI-specific Data Augmentation to Enhance the Segmentation of Right Ventricle in Multi-disease, Multi-center and Multi-view Cardiac MRI.

Accurate segmentation of right ventricle (RV) from cardiac MRI is essential to evaluate the structure and function of the RV and to further study cardiac disorders. However, it is a difficult task due to its complex crescent shape and the presence of wall irregularities in its cavity. As part of the multi-disease, multi-center, and multi-view RV segmentation in cardiac MRI challenge (M&Ms-2), we propose to solve the problem using a fully automatic deep learning method that employs different data augmentation techniques. More specifically, we applied MRI-specific based, intensity and spatial data augmentation techniques to reduce the variation among the multi-center images with various cardiac pathologies. MRI-specific data augmentation are transformations that simulate image artifacts specific to MRI such as random bias field, random ghosting and random motion artifacts. We evaluate the proposed method in the validation set of the challenge. Among the data augmentation techniques applied, the MRI-specific based data augmentation enhanced the segmentation results of both long-axis and short-axis images in terms of Dice coefficient and Hausdorff Distance (HD). From the experiments, it shows us that the usage of MRI-specific transformations alongside intensity and spatial transformations in cardiac MRI can increase the variety of the training dataset and further help to improve the generalization capabilities of the models in multi-center, multi-disease cardiac MRI images. The proposed method ranked second at the M&Ms-2 challenge.