Abstract
PURPOSE: To apply deep convolutional neural networks (CNN) to the leftventricular segmentation task in myocardial arterial spin labeled (ASL)perfusion imaging. To develop methods that measure uncertainty and that adaptsegmentation based on a specific false positive vs. false negative tradeoff. METHODS: We utilized a modified UNET architecture with Monte Carlo (MC)dropout. The model was trained on data from 22 subjects and tested on data from6 heart transplant recipients. Manual segmentation and quantitative myocardialblood flow (MBF) maps were available for comparison. We consider two globalscores of segmentation uncertainty, "Dice Uncertainty" and "MC Uncertainty",which were calculated with and without the use of manual segmentation,respectively. Tversky loss function with a hyperparameter $\beta$ was used toadapt the model to a specific false positive vs. false negative tradeoff. RESULTS: The modified UNET model achieved Dice coefficient of mean(std) =$0.91(0.04)$ on the test set. MBF measured using automatic segmentation washighly correlated to that measured using the manual segmentation ($R^2 =0.96$). Dice Uncertainty and MC Uncertainty were in good agreement ($R^2 =0.64$). As $\beta$ increased, the false positive rate systematically decreasedand false negative rate systematically increased. CONCLUSION: We demonstrate the feasibility of using CNN for automaticsegmentation of the left ventricle in myocardial ASL data. This is aparticularly challenging application because of low and inconsistentblood-myocardium contrast-to-noise ratio. We also demonstrate novel methodsthat measure uncertainty and adapt to a desired tradeoff between false positiveand false negative rates.