A Data-augmentation Framework For Contrast-agnostic Deep-learning Segmentation Of Free-running 4D CMR Under Domain Shift

Background: Free-running (FR) whole-heart imaging methods can simplify cardiac MRI workflow by eliminating the need for ECG gating, breath-holding, and extensive slice planning.¹ These methods yield comprehensive 5D datasets (3D isotropic across cardiac/respiratory phases), reformattable to any orientation retrospectively. Deep learning (DL) techniques are essential to analyze the resulting large volume of data and already enable accurate automatic segmentation of 4D FR whole-heart images.² However, domain shift degrades segmentation performance when the test data have a different contrast (e.g. pulse sequence, contrast agent, or field strength) from the training data. We proposed a data-augmentation framework to enable our DL model to generalize to unseen domains.

Methods: In this study, we consider the target domain (D₁) for out-of-domain evaluation as contrast-free 3D radial bSSFP FR images acquired at 1.5T (Sola, Siemens) in 15 participants.³ The source domain (D₂) is 3D radial GRE FR images acquired at 3T (PrismaFit, Siemens) post-gadobutrol in 20 participants.⁴ Semi-manual segmentations of the ventricular cavities and myocardium were available on these 4D FR images at end-systole and end-diastole.² We propose a data augmentation strategy (ExtAug) that widens the D₂ distribution without accessing D₁ images or labels by: (i) expanding segmentation to large extra-cardiac regions (e.g., aorta, liver) (ii) applying region-wise random intensity scaling to emulate cross-protocol contrast variability. We compared the proposed framework (nnU-Net + ExtAug trained only on D₂) to a source-only baseline (nnU-Net trained on D₂), an upper bound (nnU-Net trained on D₁), and unsupervised domain-adaptation (DA) models (SIFA⁵ and VarDA⁶) that beneficially incorporate unlabeled D₁ data during training. Evaluation used the Dice similarity coefficient (DSC), and Bland–Altman and intraclass correlation coefficient for left/right end-diastolic and end-systolic volumes and ejection fraction (EF).

Results: On the D₁ test set, the source-only model failed to adapt (low DSC), and DA methods provided limited gains despite higher compute (Table 1). In contrast, the proposed method achieved high DSC and approached the D₁-trained upper bound. Qualitatively, segmentations were coherent and close to ground truth (Fig. 1). Analyses showed high agreement for end-diastolic and end-systolic volumes, and negligible EF bias with acceptable limits of agreement (Fig. 2).

Conclusion: We validated a data-augmentation–only training strategy to mitigate cross-protocol domain shift in free-running CMR segmentation without target-domain data. Compared with source-only and DA methods, the proposed approach produced more accurate segmentations and maintained clinically acceptable agreement for ventricular volumes and ejection fraction. These results support data augmentation as a practical alternative to domain adaptation for deploying FR CMR automatic segmentation across differing scanners, sequences, and contrast conditions.

Table 1. Dice similarity coefficients (mean ± standard deviation) for left-ventricular blood pool (LVB) and myocardium (LVM), and right-ventricular blood pool (RVB) between the DL-based automatic segmentations and the ground-truth segmentation (D1-FRM) on the target domain D₁. Models include an upper bound (nnU-Net trained on D₁), a source-only baseline (nnU-Net trained on D₂), two unsupervised domain-adaptation methods (SIFA, VarDA) that beneficially incorporate unlabeled D₁ during training, and the proposed nnU-Net + ExtAug trained only on D₂ with extensive data augmentation. Training-data availability (images “anat” and segmentation “label”) is indicated. ExtAug improves over the source-only baseline and approaches the upper-bound performance.

Method	Training data				Testing against D1-FRM data
	D1 (N=15)		D2 (N=20)		LVB	LVM	RVB
	anat	label	anat	label
D1-FRA	✓	✓	×	×	0.94 ± 0.01	0.83 ± 0.01	0.92 ± 0.02
D2-FRA	×	×	✓	✓	0.63 ± 0.15	0.43 ± 0.16	0.32 ± 0.26
SIFA	✓	×	✓	✓	0.60 ± 0.23	0.47 ± 0.15	0.66 ± 0.25
VarDA	✓	×	✓	✓	0.70 ± 0.18	0.54 ± 0.18	0.56 ± 0.30
D2-FRA + ExtAug	×	×	✓	✓	0.91 ± 0.02	0.76 ± 0.06	0.88 ± 0.05

Figure 1. Qualitative results on the target domain D₁. Short-axis and pseudo four-chamber views of FR images are shown for one subject at end-diastole (left) and for another subject at end-systole (right). For each panel, the first row shows the anatomical image, the second row shows the anatomical image with ground-truth segmentation (D1-FRM), and the subsequent rows show overlays for D1-FRA (upper bound), D2-FRA (source-only), SIFA, VarDA, and D2-FRA + ExtAug (proposed). The proposed approach yields segmentations visually similar to the upper bound and ground truth segmentations with coherent delineation.
Figure 2. Clinical metric comparison through volumetric and functional agreement. Bland–Altman analyses compare ground-truth segmentations on D₁ (D1-FRM) and D1-FRA (upper bound), as well as D1-FRM and D2-FRA + ExtAug (proposed) for left and right ventricular end-diastolic volume, end-systolic volume, and ejection fraction. The x-axis shows the mean of the two methods being compared, and the y-axis shows their difference. The solid line indicates the bias; dashed lines indicate the limits of agreement (bias ± 1.96 SD). For each comparison, the intraclass correlation coefficient (ICC), standard error of measurement (SEM), and p-value are reported.