Cardiac MRI Left Ventricular Segmentation and Function Quantification Using Pre-trained Neural Networks

Abstract

Deep learning has demonstrated promise for cardiac magnetic resonance image (MRI) segmentation. However, the performance is degraded when a trained model is applied to previously unseen datasets. In this work, we developed a way to employ a pre-trained model to segment the left ventricle (LV) and quantify LV indices in a new dataset. We trained a U-net with Monte-Carlo dropout on 45 cine MR images and applied the model to 10 subjects from the ACDC dataset. The initial segmentation was refined using a continuous kernel-cut algorithm and the refined segmentation was used to fine-tune the pre-trained U-net for 10 min. This process was iterated several times until convergence and the updated model was used to segment the remaining 90 patients in the ACDC dataset. For the test dataset, we achieved Dice-similarity-coefficient of 0.81 ± 0.12 for LV myocardium and 0.90 ± 0.09 for LV cavity. Algorithm LV indices were strongly correlated with manual results (r = 0.86–0.99, p < 0.0001) with marginal biases of –8.8 g for LV myocardial mass, –0.9 ml for LV end-diastolic volume, –0.2 ml for LV end-systolic volume, –0.7 ml for LV stroke volume, and –0.6% for LV ejection fraction. The proposed approach required 12 min for fine-tuning without requiring manual annotations of the new datasets and 1 s to segment a new image. These results suggest that the developed approach is effective in segmenting a previously unseen cardiac MRI dataset and quantifying LV indices without requiring manual segmentation of the new dataset.