Cascaded Convolutional Neural Network for Automatic Myocardial Infarction Segmentation from Delayed-Enhancement Cardiac MRI

Yichi Zhang

Abstract

Automating the segmentation of myocardial infarction (MI) and no-reflow areas from Delayed-Enhancement Cardiac MRI (DE-MRI) is a critical yet challenging task due to data anisotropy and complex pathology. This paper presents a winning cascaded deep learning architecture developed for the MICCAI 2020 EMIDEC challenge. By integrating a 2D U-Net for high-resolution intra-slice detail with a 3D U-Net for volumetric refinement, the proposed model effectively leverages spatial priors to achieve state-of-the-art performance in heart damage mapping.

Introduction

Delayed-Enhancement Cardiac MRI (DE-MRI) is the gold standard for assessing myocardial viability following a heart attack. It allows clinicians to distinguish between healthy myocardium, infarcted tissue (scarring), and "no-reflow" zones where microvascular obstruction occurs. However, manual segmentation of these regions is labor-intensive and subject to significant inter-observer variability.

The primary technical challenge in automating this process is the high degree of anisotropy common in cardiac MRI protocols. While in-plane resolution is typically high (e.g., ~1.6mm), the slice thickness and spacing are often much larger (e.g., 10mm). Standard 3D networks struggle with this "stretched" data, while 2D networks fail to capture the necessary 3D anatomical context required for consistent volumetric segmentation.

Methodology

The researchers from Beihang University proposed a Cascaded Convolutional Neural Network designed to mitigate the effects of anisotropy by combining 2D and 3D processing stages.

The cascaded architecture combines a 2D U-Net for initial segmentation and a 3D U-Net for volumetric refinement.

The architecture operates in two distinct stages:

  1. Stage 1: Preliminary 2D Segmentation. A 2D U-Net processes individual slices to capture fine-grained details within the high-resolution plane. Because the resolution is highest within the slice, the 2D network is excellent at picking up intra-slice details.
  2. Stage 2: 3D Refinement. The "coarse" segmentation masks from the first stage are concatenated with the original 3D input volume. This combined input is fed into a 3D U-Net. Here, the 2D output serves as a "spatial prior," guiding the 3D network to refine the segmentation and ensure anatomical consistency across the sparsely sampled z-axis. This stage helps remove "scattered" noise that is anatomically impossible.

The framework was implemented using the nnU-Net framework and optimized using a weighted combination of Cross-Entropy and Dice loss to address the class imbalance inherent in segmenting small pathological regions.

Experimental Setup

The model was evaluated using the MICCAI 2020 EMIDEC challenge dataset, which consists of 150 patient cases (100 for training, 50 for testing). The dataset includes a mix of healthy subjects and those with myocardial pathology. Annotations cover four primary classes: the blood cavity, normal myocardium, myocardial infarction (MI), and no-reflow areas.

Results and Discussion

The cascaded approach secured the top position in the EMIDEC segmentation challenge. Quantitative results demonstrated high accuracy across all targets, with the final Dice scores reaching 0.8786 for the myocardium, 0.7124 for the infarction area, and 0.7851 for the no-reflow zone.

Visual comparison between ground truth annotations and the model's predictions on DE-MRI slices.

A key finding was the significant performance boost provided by the 3D refinement stage. For the infarction area, the Dice score improved from an average of approximately 52% in the 2D stage to over 71% in the final output. This underscores the importance of volumetric context in identifying complex, often fragmented, pathological tissues. The use of the 2D output as a spatial prior allowed the 3D network to focus on correcting errors rather than learning the entire anatomy from scratch.

Conclusion

The cascaded 2D/3D architecture effectively addresses the challenges of anisotropic cardiac MRI data. While the model achieved state-of-the-art results, segmenting pathological regions remains more difficult than segmenting healthy muscle. Future research will focus on improving the segmentation of small, irregular infarcts and evaluating the model's robustness across different MRI scanner vendors to facilitate clinical adoption.

Summary updated

BibTeX
@inproceedings{zhang2020cascaded,
  author = {Yichi Zhang},
  title = {{Cascaded Convolutional Neural Network for Automatic Myocardial Infarction Segmentation from Delayed-Enhancement Cardiac MRI}},
  booktitle = {M\&Ms and EMIDEC/STACOM @ MICCAI},
  year = {2020},
  doi = {10.1007/978-3-030-68107-4\_33},
}
Figure