TL;DR
This paper introduces a deep learning framework that leverages CNNs to extract cardiac motion from cine MRI, enabling the creation of patient-specific, dynamic 3D geometric models of the left ventricle. This advancement promises to improve diagnosis and treatment planning for cardiovascular diseases by providing more accurate and personalized cardiac models.
What Problem Does This Paper Address?
Cardiovascular diseases (CVDs) are a leading cause of morbidity and mortality. Accurate assessment of cardiac function, including wall motion and ventricular volumes, is crucial for diagnosis, prognosis, and treatment. While patient-specific geometric models derived from cardiac imaging hold great potential for detailed analysis, their construction, especially capturing dynamic motion throughout the cardiac cycle, has been a challenge. Traditional methods often involve complex segmentation and registration pipelines. This paper aims to streamline and improve the accuracy of generating these dynamic, patient-specific left ventricle (LV) myocardial models by employing a deep learning approach for motion extraction.
Methods
The core of this work is a CNN-based framework for deformable image registration. Specifically, they leverage the VoxelMorph [1] framework, a deep learning approach for estimating optical flow between image frames. This CNN is trained to predict the deformation field that maps one cardiac phase to another.
The pipeline involves several key steps:
- Image Preprocessing: Correction for slice misalignments in cine cardiac MR images using a modified U-Net model for segmentation.
- CNN-based Deformable Registration: The VoxelMorph-based CNN estimates the deformation field between cardiac frames. The loss function used incorporates both similarity (Mean Squared Error) and a smoothing term (Laplacian operator) to ensure spatially smooth registration fields.
- Mesh Generation and Propagation:
- Isosurface Meshes: The end-diastole (ED) frame's segmentation map is used to generate an initial isosurface mesh. This mesh is then propagated to subsequent frames using the deformation fields predicted by the CNN.
- Volume Meshes: Initial ED volume meshes are generated using Tetgen [17] from the surface meshes. Two methods are used to generate dynamic volume meshes:
- Direct CNN Propagation: The ED volume mesh is directly propagated using the CNN-estimated deformation fields.
- LBWARP Mesh Warping: The ED volume mesh is warped to match the dynamically propagated surface meshes using the Log Barrier-based Mesh Warping (LBWARP) method [16].
The authors also compare their CNN-based registration method against a traditional B-spline free-form deformation (FFD) non-rigid image registration algorithm implemented in SimpleElastix [9,12].
Datasets
The study utilizes the 2017 ACDC dataset [4], which consists of cine cardiac MR images from 150 subjects across five subgroups: normal, myocardial infarction (MINF), dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), and abnormal right ventricle (RV). The dataset was acquired using 1.5 T and 3.0 T MRI scanners. The authors used 110 subjects for training, 10 for validation, and 30 for testing.
Key Insights and Ideas
- Deep Learning for Motion Extraction: The paper demonstrates that deep learning, specifically CNNs like VoxelMorph, can effectively capture the complex, dynamic motion of the LV myocardium from cine MRI. This offers a more automated and potentially more accurate alternative to traditional registration methods.
- Patient-Specific Dynamic Models: The ability to generate deformable geometric models that evolve throughout the cardiac cycle is a significant step towards truly patient-specific simulations for diagnosis and treatment planning.
- Isosurface vs. Volume Mesh Propagation: The study explores two ways to generate dynamic volume meshes: direct propagation of the ED volume mesh and warping based on propagated surface meshes. The findings suggest that both methods yield comparable results, with the LBWARP method maintaining mesh topology and correspondence.
- Comparison with Traditional Methods: The CNN-based approach shows promising results when compared to traditional FFD registration, often achieving better accuracy in terms of Dice scores and mean absolute distance.
Key Results
- The CNN-based propagated isosurfaces showed a lower Mean Absolute Distance (MAD) compared to FFD-propagated isosurfaces when compared against the "silver standard" (directly segmented) isosurfaces. For instance, the CNN-Segmentation achieved MADs of 1.15 mm for Normal and 0.87 mm for DCM, while FFD-Segmentation had MADs of 1.53 mm and 1.09 mm respectively.
- The mean node-to-node distance between the CNN-propagated and LBWARP-generated volume meshes was consistently within 2 mm across all cardiac frames and pathologies, indicating good agreement between the two volume mesh generation strategies.
- The paper highlights that sequential registration (frame-to-frame) can accumulate errors over the cardiac cycle, leading to a deterioration in Dice score and MAD when propagating ED meshes to later frames. The CNN-based approach, by directly estimating larger deformations, appears to mitigate this issue.
Limitations and Future Work
- Mesh Quality at Systole: The authors acknowledge that mesh quality, particularly the scaled Jacobian values, can be lower for the end-systole (ES) phase compared to end-diastole (ED). This suggests that mesh degeneration or tangling might occur during the most dynamic phases of the cardiac cycle.
- Diffeomorphic Deformations: To further improve mesh quality and prevent tangling, the authors suggest future work focused on incorporating diffeomorphic deformations into their VoxelMorph-based method. This would ensure that the registration preserves the topological integrity of the meshes.
- Extension to RV: While the current work focuses on the LV, the authors state that the proposed framework is not limited to any pathology and can be extended to LV and RV blood-pool geometry. This opens avenues for modeling the entire heart.
Summary updated