Myocardial Infarction Segmentation in Late Gadolinium Enhanced MRI Images using Data Augmentation and Chaining Multiple U-Net

Finding the appropriate set of features in cardiac MRI images to localize different areas and anomalies of heart is an essential problem. Convolutional neural networks are known to be well suited for the task to extract features from gray scale images where the intensity is enhanced. This paper proposes a convolutional neural network architecture which can be used to localize the myocardial infarction on the heart using low resolution late gadolinium enhanced (LGE) images. LGE images are T1 weighted MRI images that use contrast agents to increase the intensity of regions where blood accumulates, as a result, areas of heart like ventricle and infarction are brighter as compared to other regions of the heart. We create a U-Net inspired model and train it with the LGE cardiac images to locate and segment the infarction. Our data set has small number of training images with low contrast, which makes it difficult to generalize the features of infarction. To tackle this issue, we propose geometric transformations and pixel intensity manipulations, that should be used for augmenting LGE images to create a diverse training data set. We also propose a chained U-Net approach to reduce the search space for segmenting infarction in LGE cardiac images. Our analysis show a reduction of error from 82% to 68% in segmentation by using the proposed augmentation and the chaining technique. However, it falls short of human level accuracy. Later part of this paper describes the limitation of our current work and lists the future work to overcome those limitations.