Healthy-to-Patients Deep learning for Automated Left Atrial Segmentation and Function analysis from Short-Axis Cine CMR (RF_TH_276)

Left atrium (LA) segmentation from Cine Short Axis (SAX) CMR is crucial for LA function assessment[1]. Yet, LA manual segmentation is subjective and time-consuming. Automated deep learning (DL) methods, though promising, face domain shift challenges when trained on diverse pathological data, leading to limited generalizability due to difficulty in obtaining data from all possible pathologies[2]. We propose a novel DL approach for domain adaptation, training solely on healthy scans and testing on unseen patient data, unlike conventional methods that mix patient data in training. Our approach automates LA segmentation across the entire cardiac cycle surpassing previous approaches restricted to 3 frames[3], [4]. We evaluated our Healthy-trained model for LA segmentation and function analysis on 3 unseen patient groups: Hypertrophic cardiomyopathy (HCM), Heart failure (HFrEF), and mitral valve regurgitation (MVR).

Methods: The study includes 118 healthy scans: 100 for training (15,150 images) and 18 for validation (2,750 images). All participants underwent multi-slice SAX cine SSFP MRI covering the LA. Only the healthy dataset was used for training/validation. To enable robust domain adaptation, we utilized a stochastic, temporally-constrained augmentation strategy within a 3D UNet architecture[5] using 2D+time stacks over the cardiac cycle(Fig.1.B-C). To improve temporal realism and consistency, composite augmentations were randomly applied in the temporal axis including Gamma correction, sharpening, rotation, flipping, cropping, & zooming. The healthy-trained model was tested on 3 diverse patient datasets (Fig.1.A). We compared performance to a state-of-the-art recent 2D-Unet baseline model for LA segmentation [6], trained with the same augmentations.


Results:

Our healthy-to-patient DL model achieved strong Dice scores of 87%, 88%, and 89%, with Hausdorff distances of 4.6mm, 6mm, & 5.7 mm for HCM, HFrEF, and MVR, significantly outperforming baseline model (Fig.2). LA functional analysis showed excellent intraclass correlation (ICC), coefficient of variation (CV), and spearman correlation for our model relative to ground truth (Fig.3B, C). Compared to the baseline model, our model showed 3-fold lower CV for LAV_max, 2-fold lower CV for LAV_min, & LAEF, with higher ICC in all parameters (Fig.3B, C).

Conclusion:

The proposed domain-adaptive LA segmentation approach, trained solely on healthy data, achieved high accuracy and generalizability across three challenging patient cohorts. By using stochastic temporally-constrained augmentation, our method achieved segmentation across all 25 cardiac cycle frames, compared to just 3 frames in previous studies, leading to accurate functional analysis against the ground truth. This work underscores the potential of training domain-adaptive models on healthy datasets, with further research needed to evaluate generalizability across multi-center and multi-vendor datasets.