Quick Fire Session 1: Basic Science and Development
QF1-010 - Fully automated CMR image segmentation in rats with myocardial infarcts using machine learning
Thursday, February 13, 2020
3:12 PM – 3:15 PM
Location: Sabal/Sago (Meeting Room in Exhibit Hall)
Background: In translational cardiovascular research, delineation of left ventricle (LV) end-systolic and end-diastolic phases is a crucial step in assessing cardiac function as requested for new therapeutic agent assessments in cardiovascular drug discovery. Recent advances in artificial intelligence (AI), such as Convolutional Neural Networks have enabled accurate segmentation of cardiac chambers from human MRI datasets . However, their application to rodent datasets, to the best of our knowledge, has not yet been explored. Here we report the results of first AI-based automated cardiac segmentation of rodent datasets. Methods: 636 end-systolic (ES) and end-diastolic (ED) short-axis datasets were manually annotated by two experts to determine grand true volumes. To achieve heterogenous dataset for model training, the images from sham-operated and rats with subacute or chronic infarcts with LV ejection fraction (EF) range of 67%-41% assessed with Segment (Medviso), were used. In addition, 144 datasets were segmented by an expert and a novice for accuracy comparison. An ensemble of volumetric fully convolutional networks was constructed based on U-net and M-net architectures [2-3]. The ensemble delineates LV contours across the heart cycle, creating time-volume curves from which end-systole and end-diastole are determined automatically. ES, ED volumes and EF derived from these curves were compared against human segmentation using Dice score and Bland-Altman analysis. Results: The automated LV segmentation was in good agreement with manual segmentation by an expert at both ED and ES phases (Fig1-2). The mean Dice score ± SD for the automated segmentation vs. expert was 0.93±0.05 (0.92±0.02 for ES and 0.95±0.01 for ED), and was higher than for expert vs. novice agreement (0.90±0.07 mean, 0.87±0.03 ES, 0.93±0.02 ED). The automated segmentation led to a higher accuracy when compared to automated-expert vs. novice-expert with the mean relative difference of 5.0±4.4 % vs. 8.2±4.6 % for EDV, 8.2±6.1% vs. 25.9±9.1% for ESV, and 6.3±2.4% vs. 6.8±2.3% for EF. The inter-reader variability assessed from repeated scans for expert and automated segmentation was similar with SD of ΔEF=2.4%. Such level of precision will enable detection power of LV functional improvement with the same animal number but 18 times faster, on the account of shorter analysis time taking ~1.6 min/animal for automatic segmentation vs. ~30 min when performed manually. Conclusion: The proposed method allows fast, accurate and fully automated analysis of cardiac chamber size and function in rodents. AI will be essential when assessing pharmacological interventions designed to improve cardiac function with increased speed and high precision.