[Paper Review] Anatomically-Informed Deep Learning on Contrast-Enhanced Cardiac MRI for Scar Segmentation and Clinical Feature Extraction

Haley G. Abramson, Dan M. Popescu+6 more|arXiv (Cornell University)|Oct 21, 2020

Advanced MRI Techniques and Applications4 citations

TL;DR

This paper presents ACSNet, a fully automated, anatomically-informed deep learning framework for left ventricle and scar/fibrosis segmentation on late gadolinium enhancement cardiac MRI (LGE-CMR). Using three cascading U-Net-based networks with synthetic 'LGE-like' cine images for data augmentation, it achieves expert-level accuracy (96% balanced accuracy for LV, 75% for scar) and clinically relevant feature extraction with minimal inter-observer variability.

ABSTRACT

Visualizing disease-induced scarring and fibrosis in the heart on cardiac magnetic resonance (CMR) imaging with contrast enhancement (LGE) is paramount in characterizing disease progression and quantifying pathophysiological substrates of arrhythmias. However, segmentation and scar/fibrosis identification from LGE-CMR is an intensive manual process prone to large inter-observer variability. Here, we present a novel fully-automated anatomically-informed deep learning solution for left ventricle (LV) and scar/fibrosis segmentation and clinical feature extraction from LGE-CMR. The technology involves three cascading convolutional neural networks that segment myocardium and scar/fibrosis from raw LGE-CMR images and constrain these segmentations within anatomical guidelines, thus facilitating seamless derivation of clinically-significant parameters. In addition to available LGE-CMR images, training used "LGE-like" synthetically enhanced cine scans. Results show excellent agreement with those of trained experts in terms of segmentation (balanced accuracy of $96\\%$ and $75\\%$ for LV and scar segmentation), clinical features ($2\\%$ difference in mean scar-to-LV wall volume fraction), and anatomical fidelity. Our segmentation technology is extendable to other computer vision medical applications and to problems requiring guidelines adherence of predicted outputs.

Motivation & Objective

• To address the high inter-observer variability and time-intensive nature of manual LGE-CMR scar and myocardial segmentation by automating the process.
• To improve segmentation robustness across diverse imaging protocols and pathologies by incorporating anatomical constraints into deep learning.
• To enable accurate, reproducible extraction of clinical features such as scar-to-LV wall volume fraction for arrhythmia risk stratification.
• To overcome data scarcity in LGE-CMR by leveraging synthetic 'LGE-like' enhancement on cine MRI to augment training.
• To ensure anatomical fidelity in deep learning outputs, enabling reliable derivation of hemodynamic and fibrotic parameters.

Proposed method

• A cascading three-network architecture: (1) left ventricle (LV) segmentation, (2) myocardium segmentation, and (3) autoencoder-based anatomical prior learning.
• Synthetic 'LGE-like' enhancement is applied to end-diastolic (ED) cine MRI using style transfer to simulate LGE contrast, expanding training data without real LGE scans.
• The LV and myocardium segmentation networks use a weighted combination of balanced cross-entropy and Tversky loss to optimize for both pixel-level and image-level segmentation accuracy.
• The Tversky loss function incorporates a beta parameter (0.6 for LV, 0.4 for myocardium) to penalize false positives more heavily, improving boundary precision.
• Anatomical constraints are enforced via an autoencoder trained on ground truth myocardial masks, learning a low-dimensional latent space (d=16, p=6) that encodes anatomical plausibility.
• Post-processing includes morphological erosion with a 3×3 ellipsoid kernel and component filtering to remove small false positives, followed by FWHM thresholding to isolate dense core scar.

Experimental results

Research questions

• RQ1 Can a fully automated deep learning framework achieve expert-level accuracy in left ventricle and scar segmentation on LGE-CMR without manual intervention?
• RQ2 Does incorporating synthetic 'LGE-like' enhancement on cine MRI improve generalization and performance across diverse imaging protocols and pathologies?
• RQ3 Can anatomical priors encoded via an autoencoder enhance segmentation fidelity and reduce anatomical outliers in deep learning outputs?
• RQ4 To what extent does the method preserve clinically relevant features such as scar-to-LV wall volume fraction compared to expert readings?
• RQ5 How does the model perform across different cardiac regions (apex, mid-ventricle, base) in terms of segmentation consistency and accuracy?

Key findings

• The method achieved a balanced accuracy of 96% for left ventricle segmentation and 75% for scar/fibrosis segmentation, demonstrating strong agreement with expert annotations.
• Scar-to-LV wall volume fraction calculated from ACSNet segmentations differed by only 2% from expert-derived values, indicating high clinical relevance.
• The model maintained consistent performance across all cardiac regions (apex, mid-ventricle, base), avoiding the high-performing-outlier problem common in average-based evaluations.
• The use of synthetic 'LGE-like' cine images significantly improved data efficiency and generalization, especially given limited real LGE-CMR datasets.
• Post-processing with morphological erosion and component filtering reduced false positives and improved segmentation fidelity without compromising true positive detection.
• The final model selected a latent space dimension of d=16 with p=6 components, balancing reconstruction accuracy and model complexity based on NLL and AIC criteria.