TL;DR
This paper introduces a novel deep learning pipeline for automated, high-throughput analysis of cardiac MRI in rat models of pulmonary hypertension (PH). The system accurately segments both ventricles and generates 3D wall motion maps, providing a crucial translational tool to bridge findings from preclinical rodent studies to human PH research.
What Problem Does This Paper Address?
Accurately assessing cardiac function and remodeling is vital for understanding diseases like pulmonary hypertension (PH). While cardiac MRI is the gold standard, manual segmentation of cardiac images is time-consuming and requires specialized expertise. This is particularly challenging in preclinical research using animal models, where large datasets are often generated for disease mechanism studies. Existing automated methods for animal cardiac MRI segmentation are limited, especially for the right ventricle (RV), which is heavily impacted in PH. This paper aims to bridge this gap by developing an automated pipeline for bi-ventricular segmentation and 3D wall motion analysis specifically tailored for rat models of PH, enabling more efficient and consistent research.
Methods
The core of the proposed solution is a deep learning pipeline that combines a fully convolutional network (FCN) for initial segmentation with an atlas-fitting approach for generating 3D cardiac models.
- Automated Segmentation: A 2D FCN was trained on a dataset of short-axis cine cardiac MRI scans. This network predicts segmentations for both the left ventricle (LV) and right ventricle (RV) cavities and walls.
- Atlas Refinement & 3D Modeling: The FCN's output from low-resolution cine images was then co-registered with high-resolution (HR) cardiac MRI atlases. This multi-atlas label fusion and subsequent rigid registration to an average template allowed for the generation of anatomically accurate 3D segmentations.
- Wall Motion Analysis: A marching cubes algorithm was used to create 3D cardiac mesh models from the segmented data. These meshes were then used to calculate regional wall motion (longitudinal, radial, and circumferential) between end-diastole and end-systole.
The pipeline was designed to be efficient, with segmentation completed in under 1 second per scan.
Datasets
The study utilized a dataset of 163 short-axis cine cardiac MRI scans collected longitudinally from:
- Healthy Sprague Dawley or Wistar Kyoto rats.
- Monocrotaline (MCT)-induced PH rat models at 2 and 4 weeks post-induction.
- Sugen-hypoxia (SuHx)-induced PH rat models at 4, 6, and 8 weeks post-induction.
These datasets were split into training, validation, and testing sets. Additionally, a separate dataset of female rat controls and PH models was used to assess the generalizability of the FCN. A subset of high-resolution (HR) segmentations was also used to build the atlases for the refinement step.
Key Insights and Ideas
- Translational Potential: The pipeline's ability to generate 3D wall motion maps in rat models that mirror patterns observed in human PH patients is a significant step towards translating preclinical findings.
- RV Focus is Crucial: The study highlights the importance of accurate RV segmentation and analysis, as the RV is disproportionately affected in PH and its function is a key predictor of survival.
- 2D FCN Efficacy: Surprisingly, a 2D FCN architecture outperformed a 2.5D network for this specific rat dataset, suggesting that simpler architectures can be highly effective with appropriate training data.
- Atlas-Based Refinement: Integrating an atlas-refinement step significantly improves the anatomical accuracy of the segmentations and enables the generation of detailed 3D cardiac meshes.
- Automated vs. Human Performance: The automated FCN demonstrated superior consistency and lower bias compared to manual segmentation, even outperforming human readers in some cardiac indices.
Key Results
- Segmentation Accuracy: The automated segmentation achieved a high Dice metric, with mean values above 0.92 for most labels (LV cavity, LV wall, RV cavity) and 0.836 for the RV myocardial wall. The model achieved this in less than 1 second per scan.
- Cardiac Parameter Agreement: Cardiac indices (EDV, ESV, SV, EF, Mass) derived from automated segmentations showed high correlation (r > 0.91) with manual segmentations, with minimal bias and narrow limits of agreement.
- RV Hypokinesia in PH: PH rats exhibited significant right ventricular hypokinesia (-37.7% ± 12.2 for MCT; -38.6% ± 6.9 for SuHx) compared to controls, primarily due to reduced basal longitudinal and apical radial motion.
- Generalizability: The FCN performed well on unseen data, including a separate dataset of female rats, demonstrating good generalizability.
- Interoperator Variability: The automated method outperformed human readers in terms of consistency and bias for most cardiac indices.
Limitations and Future Work
- Image Quality Dependence: The study acknowledges that severe ventricular remodeling, particularly near end-stage heart failure, can impact image quality and lead to segmentation inaccuracies. A visual quality check is recommended for such cases.
- Scanner Variability: The pipeline was trained on data from a specific imaging center. Implementing it on data from different scanners might require visual quality control and fine-tuning via transfer learning.
- Limited Timepoints for Atlas: The atlas pool was limited to specific timepoints. Expanding this pool could further enhance robustness.
- Future Work: The authors suggest that the pipeline could be extended to study other cardiac conditions impacting the RV or LV, and that the methodology aligns with broader trends in clinical cardiac imaging AI development, potentially stimulating further research into similar assistive methods.
Summary updated