Deep Generative Model based Virtual LGE and Feasibility Evaluation in Hypertrophic Cardiomyopathy (RF_FR_375)

Late gadolinium enhancement (LGE) is a standard clinical method that relies on contrast agents and has long imaging time. In contrast, cine is a more widely used sequence, holds promise as a contrast agent-free technique for detecting myocardial abnormalities observed in LGE. In this study, we propose a Registration-based Generative Adversarial Network-Convolutional Block Attention Module (RegGAN-CBAM) Model to generate virtual LGE and evaluate the effectiveness of its diagnostic accuracy for hypertrophic cardiomyopathy (HCM).

Methods:

Study population and image acquisition

This retrospective study received approval from the local institutional review board. All participants provided informed consent. A total of 149 HCM patients (age (years): 50 ±16 (mean ± standard deviation), gender (male/female): 87/62) were included. CMR was conducted using a 3T MRI machine (Ingenia, Philips Healthcare, Best, The Netherlands) with a DS torso coil.

RegGAN-CBAM model

The model architecture is shown in Figure 1. The RegGAN was utilized to synthesis LGE and to refine the generated results, while the CBAM employs attention mechanisms through two modules: 1) channel attention enhances feature learning; 2) spatial attention facilitates the acquisition of spatially information. The registration module operates as a label noise model to enhance the generated outcomes. Furthermore, the datasets were randomized into 3 independent groups for training (60%), validation (10%), and testing (30%). The model training parameters were configured with epoch of 80, 150, 200 and learning rate of 0.0001.

Evaluation and analysis

Statistical analysis was performed to quantify myocardial enhancements employing native and virtual LGE. Spearman's correlation test was utilized to evaluate the correlation of myocardial enhancements within myocardium regions. A Bland-Altman plot was generated to assess the agreement in myocardial enhancements between native and virtual LGE. Moreover, box plots were generated to evaluate the difference in myocardial enhancements between the native and virtual LGE, employing a paired t-test.

Results:

Intermediate results during training

Figure 2a demonstrates the evolution of virtual LGE generation during training of RegGAN-CBAM at epochs 80, 150, and 200. It is evident that as the iterations progress, the virtual LGE becomes increasingly clearer.

Image quality

Typical images of cine, virtual LGE (epoch = 200), and native LGE are shown in Figure 2b. Visually, the virtual LGE exhibit similarity to the native LGE, particularly in texture and edge characteristics.

Statistical analysis

The myocardial enhancements were determined through the delineation of regions of interest (ROIs) on the myocardium LGE images (native and virtual LGE, as shown in Figure 3a). The correlation between native and virtual LGE myocardial enhancements (r = 0.908, p < 0.01) is significant (Figure 3b). Analysis of the Bland-Altman plot reveals that most data points cluster around the mean difference, with the absence of outliers, indicating a high consistency in myocardial enhancements between native and virtual LGE (Figure 3c). Figure 3d illustrates that there is no significant difference between these two groups, with a p-value of 0.847.

Conclusion:

This study illustrates the feasibility of RegGAN-CBAM in synthesizing LGE. The virtual LGE displays similarity to native LGE in imaging characteristics, and their myocardium enhancement values exhibit strong correlation and consistency.