Deep-CodecG*: A Generalized Deep Autoencoder for Robust Segmentation of Left Atrium in Cardiac MRIs Deep-CodecG* for Robust Segmentation of Left Atrium.

Abstract

The left atrium receives oxygenated blood from pulmonary veins and is a vital organ concerning
congestive heart failure. Several deep learning-based architectures and learning methodologies have
been proposed for left atrium semantic segmentation. These studies have shown good performance in
learning known datasets. However, generalization remains challenging. In this work, we propose a deep
auto-encoder architecture with generalization ability which we call Deep-CodecG*. The proposed model
utilized a CNN-based auto-encoder in which the standard convolution is replaced with a two-convolution
layer block. This proposed model is generalization enabled with a proper parameterization for (near-)
optimal performance. The proposed Deep-CodecG* improves performance on unseen test data, a dice
score of 0.95, which is 6.3% higher than that of a standard auto-encoder. The proposed model gave higher
sensitivity, specificity, Jaccard, and structural similarity values and lower Hausdorff distance indicating
improvement over an autoencoder with similar two-convolution layer blocks. Though these quantitative
improvements seem marginal, they are shown to have a significant impact. The segmented left atrium
images match the ground-truth data very closely. Thus, the proposed Deep-CodecG* architecture for left
atrium segmentation exhibits well-generalized and robust performance over various image datasets