Abstract 16017: Anatomically-Guided Deep Learning (DL) Approach to Late Gadolinium Enhanced (LGE)-CMR Left Ventricle (LV) Segmentation Enables Efficient and Accurate Clinical Analyses

Abstract

Rationale: LV segmentation is crucial to analyzing heart pathologies. DL applications for LGE-CMR image segmentation offer automation and reliability over costly feature learning and manual input. However, existing DL algorithms suffer from poor quality and quantity of training data, lack of clinical relevance in evaluation methods, and low generalizability. Objective: We propose a neural network approach to LV segmentation with superior performance and ensured anatomical fidelity on various imaging protocols and patient populations, with direct clinical evaluation via LV volume calculation and ventricular arrhythmia (VA) risk assessment. Methods: We train a multistage network for LV segmentation on 4,000 LGE-CMR (PROSe-ICD study) and cine (MICCAI 2009 and 2017 datasets) short-axis images, where cine scans augment the data via a cine-to-LGE style transfer. At testing, images enter a network that crops and refines the region-of-interest (fig. 1A), which is fed into a network for differentiation of viable and hyper-enhanced myocardium (fig. 1B). An autoencoder network takes the union of these images, encodes them to a densely populated, reduced space, finds the nearest neighbor, and decodes them to anatomically correct segmented scans (fig. 1C). To assess VA risk, we fit a Cox regression model to survival data with image features (ex. LV volume) from manual or predicted segmentations. Results: Our segmented LGE-CMR scans (0.81 Dice score) surpass an inter-observer Dice score of 0.76, upholding anatomical guidelines across ambiguous regions such as apex and base (fig. 1D). Differences in calculated LV volumes (accuracy 98.4%, precision 78.9%) and C-indices of Cox regression models that stratify VA risk using automatic versus manual segmentations (0.62 vs. 0.60) show no statistical significance. Conclusion: We developed a DL LV segmentation algorithm tailored to clinical application, ensuring relevant predictions and outperforming inter-observed Dice scores.