Abstract
Myocardial pathology segmentation (MyoPS) can be a prerequisite for the accurate diagnosis and treatment planning of myocardial infarction. However, achieving this segmentation is challenging, mainly due to the inadequate and indistinct information from an image. In this work, we develop an end-to-end deep neural network, referred to as MyoPS-Net, to flexibly combine five-sequence cardiac magnetic resonance (CMR) images for MyoPS. To extract precise and adequate information, we design an effective yet flexible architecture to extract and fuse cross-modal features. This architecture can tackle different numbers of CMR images and complex combinations of modalities, with output branches targeting specific pathologies. To impose anatomical knowledge on the segmentation results, we first propose a module to regularize myocardium consistency and localize the pathologies, and then introduce an inclusiveness loss to utilize relations between myocardial scars and edema. We evaluated the proposed MyoPS-Net on two datasets, i.e., a private one consisting of 50 paired multi-sequence CMR images and a public one from MICCAI2020 MyoPS Challenge. Experimental results showed that MyoPS-Net could achieve state-of-the-art performance in various scenarios. Results proved the superiority and generalizability of MyoPS-Net, and more importantly, indicated a practical clinical application.
Introduction
Accurately identifying myocardial scars and edema is critical for diagnosing and treating myocardial infarction (heart attack). Clinicians typically rely on several Cardiac Magnetic Resonance (CMR) sequences to visualize different pathological features: Late Gadolinium Enhancement (LGE) is the gold standard for identifying scars, while T2-weighted images highlight edema.
However, two major hurdles exist in current workflows. First, effectively fusing data from five different CMR sequences (bSSFP, LGE, T2, T1 mapping, and T2* mapping) is difficult due to varying appearances and noise levels. Second, clinical reality often involves missing data; for instance, patients with renal issues cannot receive the contrast agents required for LGE scans. MyoPS-Net addresses these challenges with a "flexible combination" architecture that maintains high accuracy even when certain imaging sequences are missing.
Methodology
The MyoPS-Net framework introduces three novel components to handle multi-sequence data and enforce anatomical constraints:
Cross-Modal Feature Fusion (CMFF)
Instead of simply stacking images at the input layer, the CMFF module uses multiple encoders that interact via pixel-wise max operations. This allows the network to dynamically extract and fuse the most salient features from whichever CMR sequences are available, ensuring robustness to missing modalities.
Myocardium Prior and Consistency (MPC) Module
Since scars and edema are anatomically confined within the heart muscle, the MPC module first localizes the myocardium. It then applies a consistency loss to ensure that pathology segmentation remains strictly within these anatomical boundaries, significantly reducing false positives in non-myocardial tissue.
Pathology Inclusiveness (PI) Loss
MyoPS-Net incorporates medical domain knowledge through a PI loss. In clinical practice, myocardial scars are almost always contained within a larger area of edema. The PI loss penalizes the model if it predicts a scar outside of an edematous region, forcing the AI to learn realistic spatial relationships.
Experimental Results
The model was evaluated on a private dataset of 50 paired multi-sequence CMR images and the public MICCAI 2020 MyoPS Challenge dataset.
- State-of-the-Art Performance: MyoPS-Net achieved Dice scores of 0.661 for scars and 0.742 for edema on the public challenge dataset, outperforming established models like nnU-Net and EfficientSeg.
- Ablation Success: Each module added measurable value. The CMFF module boosted the scar Dice score from 0.560 to 0.614, while the addition of MPC and PI modules further increased it to 0.656.
- Clinical Flexibility: The authors demonstrated that a unified model (MyoPS-Net-mix) can generalize across different hospital protocols. Even when LGE sequences were removed, the model leveraged mapping sequences (T1 and T2*) to produce clinically relevant segmentations, proving its utility for contrast-free imaging.
Discussion and Conclusion
MyoPS-Net represents a shift from pure computer vision to physically constrained medical imaging. By embedding anatomical rules—such as the spatial relationship between scars and edema—the researchers significantly improved segmentation precision. While the current model assumes pre-registered images, future work aims to integrate registration and segmentation into a single framework. MyoPS-Net’s ability to handle complex combinations of CMR sequences makes it a highly practical tool for real-world clinical environments where data may be incomplete.
Summary updated