Automated Left Atrial Segmentation from MRI Sequences

This paper presents a novel algorithm for automated left atrial segmentation, from cine MRI sequences, using a modified deep CNN model. Segmentation of the LA is necessary for a variety of medical and similar applications. The analysis of the LA often involves manual tracing of the boundary of the chamber, which is subject to human errors, and is a complex and time-consuming process. An automated and precise LA segmentation model is thus quite desirable in society. The objective is to build a neural network based on the pre-trained CNN Inception V4 architecture and to predict a compressed vector by applying a multi-layer autoencoder, which is then to be back- projected into the segmentation contour of the LA to perform delineation using open contours. Quantitative evaluations are performed to compare the proposed method with the current state-of-the-art U-net method. MR images are made to undergo Late Gadolinium Enhancement before it is used as input to the CNN model, improving the quality of the image. The CNN architecture was enhanced by adding convolution and ReLU layers which help in object-background separation. The model was trained using images from 1000 MRI patients.