Interpretable machine learning for automated left ventricular scar quantification in hypertrophic cardiomyopathy...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Scar quantification on cardiovascular magnetic resonance (CMR) late gadolinium enhancement (LGE) images is crucial for risk stratifying hypertrophic cardiomyopathy (HCM) patients.
• Scar burden is a key predictor of clinical outcomes in HCM.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for automated contouring of left ventricular (LV) borders and scar quantification on CMR LGE images in HCM patients.
• To assess the accuracy and generalizability of the ML model.

Main Methods:

• A 2-dimensional convolutional neural network (CNN) was trained on 2557 LGE images from 307 HCM patients.
• The model was trained using 80% of the data and tested on the remaining 20%, with manual segmentation by two experts serving as the gold standard.
• Performance was evaluated using Dice Similarity Coefficient (DSC), Bland-Altman analysis, and Pearson's correlation.

Main Results:

• The ML model achieved good to excellent DSC scores for LV endocardium (0.91 ± 0.04), epicardium (0.83 ± 0.03), and scar segmentation (0.64 ± 0.09).
• The model demonstrated low bias and limits of agreement (-0.53 ± 2.71%) and high correlation (r = 0.92) for LGE to LV mass percentage.
• The automated algorithm requires no manual image pre-processing and was trained on data from multiple experts and software, enhancing its generalizability.

Conclusions:

• A fully automated, interpretable ML algorithm enables rapid and accurate scar quantification from CMR LGE images in HCM patients.
• This approach facilitates improved risk stratification and clinical outcome prediction.