Fully Automatic Left Atrial Strain Quantification via Multi-task Learning on Cardiac Cine MRI.
TL;DR
A fully automatic, multi-task learning framework for LA strain quantification, validated in multi-center two-vendor data, enhances tracking accuracy and speed over prior methods, enabling rapid, scalable atrial function assessment in routine care.
Abstract
Background
Cardiac magnetic resonance feature tracking (CMR-FT) of left atrial (LA) strain is hindered by thin-wall contouring errors, motion heterogeneity, and temporal drift, while manual or landmark-based methods lack reproducibility and scalability.
Methods
We retrospectively collected a multi-center, two-vendor cine MRI dataset. A multi-task learning model was developed to quantify LA strain directly from two-chamber and four-chamber cine images, by coupling a groupwise registration network with a segmentation network through a spatiotemporal cross-attention module and synergistic losses. Performance was benchmarked against common feature tracking algorithms, including optical flow, pairwise registration, and VoxelMorph, via various metrics such as mean-squared error, contour distance, mitral annular tracking accuracy, and drift error. Diagnostic performance to distinguish healthy from diseased subjects was assessed by ROC analysis.
Results
546 subjects (142 healthy; age 49±18 years; 343 male) were included for method development and internal/external testing. The proposed method outperformed all other methods in tracking accuracy and reduced the drift effect commonly observed in optical flow and pairwise registration to a level comparable to fixed-reference learning-based registration. Inference required half a second. Automatic strains agreed closely with manual-segmentation-derived values (reservoir r=0.95, conduit r=0.96, booster r=0.92; all p<0.001). In the external dataset, all three strain components were lower in diseased subjects than normal controls (reservoir 24.4±13.2% vs 45.7±12.6%, conduit 14.1±8.8% vs 31.1±10.2%, and booster 10.3±6.4% vs 14.6±5.1%, all p<0.001). Compared with alternative methods, the automatic reservoir strain achieved the highest discriminative power across multiple diseased groups (AUC: 0.81-0.97).
Conclusion
A fully automatic, multi-task learning framework for LA strain quantification, validated in multi-center two-vendor data, enhances tracking accuracy and speed over prior methods, enabling rapid, scalable atrial function assessment in routine care. Source code is available at SJTU-CMRLab/Dual_Task_LA_Strain_Quantification.