Development of a Novel Deep Learning System for Aortic Valve Flow Quantification on Cardiac Magnetic Resonance Imaging.
Alex Bratt, M.D.1, Meredith Pollie, B.S.2, Noel Codella, Ph.D.3, Wayne Colizza, M.D.1, Rocio Perez-Johnston, M.D.4, Jiwon Kim, M.D.2, Jonathan Weinsaft, M.D.2.
1Weill Cornell Medicine/New York Presbyterian Hospital, New York, NY, USA, 2Weill Cornell Medicine, New York, NY, USA, 3IBM TJ Watson Research Center, Yorktown Heights, NY, USA, 4Memorial Sloan Kettering Cancer Center, New York, NY, USA.
OBJECTIVE: To test a novel machine learning derived cardiac magnetic resonance (CMR) algorithm for aortic valve flow quantification. CMR is increasingly used as a quantitative reference for valvular flow, but current clinical application employs manual segmentation - which can be subjective and time consuming. Machine learning has been employed for general cardiac image analysis, but application for flow quantification has been limited.
METHODS: Aortic phase contrast (PC) CMR datasets were collected to develop and validate a deep learning algorithm. PC-CMR was acquired using commercial scanners (typical temporal resolution 30msec, venc 1.5m/sec). Manual segmentation was used to classify pixels as either aortic valve or non-valve, establishing ground-truth segmentation maps. The machine learning model employed deep neural network based on the popular U-net architecture, modified by integrating residual modules, which improve gradient propagation during training and enable creation of deeper networks. Model performance was evaluated using the Dice coefficient, Jaccard index, and net forward volume between manual and model-generated segmentations. Patients with prosthetic aortic valves or congenital heart disease were excluded.
RESULTS: 153 patients undergoing CMR aortic valve flow quantification were studied. The model was trained in an initial derivation cohort (n=128) using a reference of manual segmentation and tested in a subsequent cohort (n=25). In all cases, automated segmentation yielded interpretive results and processing time was rapid (~0.75 seconds/case). Mean Dice coefficient and Jaccard index per case over the test set were 0.94 (CI 0.94-0.95) and 0.89 (CI 0.88-0.90), respectively, consistent with high agreement vs. manual segmentation. Regarding actual flow, automated and manual segmentation yielded near exact agreement (mean ∆0.7 mL [CI 0.43-0.97mL]; p=0.46). Figure provides a representative example of machine learning derived aortic valve flow quantification produced by the neural network algorithm.
CONCLUSIONS: This study provides initial validation of a fast and accurate deep learning method for automatic aortic valve segmentation on CMR. Future work will test the machine learning flow quantification algorithm in larger cohorts, including quantification of aortic and mitral regurgitation.
Back to 2018 Program