Clinical Evaluation of A-LIKNet

Objectives: To evaluate the diagnostic accuracy and quantitative agreement of A-LIKNet (attention-incorporated network for sharing low-rank, image, and k-space information) deep learning (DL)-accelerated 2D cardiac CINE MRI acquired in a single breath-hold, compared with standard multi-breath-hold CINE sequences, for assessing biventricular volumes and function. Materials and Methods: In this single-center study, A-LIKNet DL-reconstructed CINE images were acquired at 3 acceleration factors (8×, 16×, and 24×) in 42 subjects using a single breath-hold protocol. Quantitative parameters, including left and right ventricular end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), ejection fraction (EF), and left ventricle myocardial mass (MM), were derived from standardized segmentation workflows. These were compared against standard multi-breath-hold CINE sequences and tested for agreement using Bland–Altman analysis and linear regression. Image quality was assessed using the coefficient of variation (CoV). Results: Excellent agreement was observed between A-LIKNet DL-accelerated and standard CINE imaging for EDV, ESV, SV, and EF, with 95% limits of agreement (LoA) falling within predefined equivalence margins for nearly all parameters. No significant proportional bias was detected. Myocardial mass showed wider variability and exceeded equivalence thresholds, likely due to less distinct epicardial borders in higher acceleration. CoV values increased with higher acceleration, reflecting mild degradation in image quality, although segmentation performance remained robust across all levels. Conclusion: A-LIKNet deep learning accelerated CINE sequences enable rapid and reliable assessment of cardiac function with excellent agreement to standard imaging. From a clinical perspective, image quality remains acceptable up to an acceleration factor of 16×, supporting routine application, while 24× acceleration may be reserved for selected use cases requiring maximal speed.