Deep learning with dual Image-quality enhancement and harmonized-calibration for improved Biventricular segmentation from Cine CMR (RF_TH_272)

Background: Variability in image quality and intensity distribution in cardiac CMR datasets can significantly hinder automated segmentation methods. Traditional enhancement techniques often struggle with complex artifacts, and synthetic augmentation may not represent realistic data distributions. To overcome these issues, we developed a deep learning approach that enhances image quality and enforces data harmonization by recalibrating datasets to align with high-quality reference images, allowing the model to focus on intrinsic features instead of variability in image quality. The aim of this study is to evaluate the performance of this technique on biventricular segmentation accuracy, and LV/RV volumes.

Methods: We used the public ACDC dataset [1], containing 1,367 2D slices and 150 3D end-diastolic volumes from 150 patients (30 healthy, 30 myocardial infarction, 30 dilated cardiomyopathy, 30 hypertrophic cardiomyopathy, and 30 with abnormal right ventricles). 
To enhance image quality, we applied a modified Deep Adversarial Structured Quality Enhancement (DASQE) network [2], employing the analysis pipeline in Fig.1.  This network, built on an autoencoder architecture with adversarial learning, enhances image quality by aligning low-quality images intensity distribution with reference high-quality targets. It employs a quality-aware perceptual loss to recalibrate the dataset, ensuring consistent intensity distribution and reducing variability. Images were categorized into target (high-quality, consistent style), source(high-quality, different style) and low-quality data. We used 20 selected 20 high-quality 2D cine images as target.  Enhanced images were segmented using a 2D UNet model in PyMic [3] for automated segmentation of the left ventricle (LV), right ventricle (RV), and LV myocardium. Performance, versus a baseline model with unenhanced images, was evaluated using 5-fold cross-validation Dice scores and assessments of agreement in EDV and myocardial mass.

Results: The proposed enhancement-calibration model visually improved data quality (Fig. 2) and significantly boosted performance across all metrics (Fig. 2b, Fig. 3). The highest improvement was seen in the RV, where Dice scores increased from 0.90 to 0.94 (p < 0.001), RV volume ICC rose from 0.65 to 0.95, and CoV decreased nearly 3-fold from 9.33 to 3.98 (Fig. 3a). The model also significantly improved LV segmentation, with Dice scores for LV myocardium rising from 0.86 to 0.89 (p = 0.01) and from 0.96 to 0.97 for LV endocardium (p < 0.001), with higher ICC and lower CoV for LV EDV and LV mass, i.e. better agreement with ground truth (Fig. 3a, b).

Conclusion: The proposed deep learning-based enhancement-calibration model significantly improved image quality and segmentation accuracy for both the RV and LV by harmonizing and recalibrating image intensity distribution. Results showed up to 3-fold better agreement with ground truth suggesting the proposed technique’s potential to boost performance and reliability of automated cardiac CMR image analysis.