Deep Learning for End-to-End Atrial Fibrillation Recurrence Estimation

Riddhish Bhalodia1, Anupama Goparaju2, Tim Sodergren1, Ross Whitaker1, Alan Morris3, Evgueni Kholmovski3, Nassir Marrouche4, Joshua Cates3, Shireen Elhabian1
1Scientific Computing and Imaging Institute, University of Utah, 2University of Utah, 3Comprehensive Arrhythmia Research & Management Center, 4Comprehensive Arrhythmia Research and Management Center


Motivation: Left atrium shape morphology has been shown to be an independent predictor of recurrence after atrial fibrillation (AF) ablation. Nonetheless, Population-level shape representations in form of image segmentations and correspondence models (standard method - SM) derived from cardiac MRI require significant human resources with sufficient anatomy-specific expertise. In this study, we propose a machine learning approach to study AF recurrence by predicting shape descriptors from raw MRI images (proposed method - PM), with NO image pre-processing involved. Methods: We train a deep network to estimate a low-dimensional shape representation, in the form of PCA loadings, directly from 3D images. With a representative set of 100 AF patients collected retrospectively, 10 PCA loadings were found to capture 90% shape variability. We used a multi-layer perceptron (MLP) to regress AF recurrence. Results: For deep network training, we augment the data by sampling from the PCA space of 75 MRI scans ('seen data'). Remaining 25 are isolated ('unseen data'). PCA loading from the SM and the ones estimated from images were found as statistically same by Hotelling T2 statistic of 6.01 with 82% confidence (99% confidence on seen and 63% confidence on unseen). The recurrence probability prediction by MLP (trained using SM PCA loadings) using PCA loadings from SM and PM, are statistically same by T-Test with 79.6\% confidence. Conclusion: Proposed method performs similar to the standard shape modeling method and produce statistically similar outcomes on AF recurrence, eliminating the need for heavy pre-processing and associated manpower. Occasional errors of the method can be attributed to the huge variability in image intensities arising from different scanners and acquisition protocols, the performance can be improved with smart data selection while training.