AI-Guided Compressed Sensing for Accelerated Cardiac Cine MRI

Background: Cardiac cine MRI is essential for evaluating cardiac anatomy and function. Conventional compressed sensing (CS)¹ techniques enable accelerated acquisition but suffer from long reconstruction times, limiting clinical applicability. Residual convolutional recurrent neural network (res-CRNN)² is utilized in cardiac cine reconstruction with ultrafast reconstruction time; however, the resulting images often show temporal blurring that reduces myocardial edge sharpness and leads to inaccurate cardiac functional measurements. These shortcomings arise from overly strong temporal regularization imposed by AI networks together with insufficient enforcement of data consistency (DC). We therefore introduce an AI-guided CS framework that leverages the rapid inference capability of AI and the strong data consistency of CS, providing sharp and physiologically reliable cardiac cine reconstructions.

Methods: We developed an AI-guided compressed sensing (ACS) reconstruction framework that augments conventional CS, as illustrated in Figure 1. The AI-reconstructed cine images were used as initialization for subsequent CS iterations, accelerating convergence and improving stability. To ensure compatibility between the AI output and CS reconstruction, a normalization step was applied by aligning A^HA(x) (with A denoting the forward encoding operator including coil sensitivity, Fourier transform, and sampling mask) with A^Hb (where b is the undersampled k-space data).
Image sharpness was further preserved by incorporating strengthened k-space data consistency constraints within each iteration. Owing to the high-quality initialization, only a few iterations were required, adding approximately 1 second per slice. The proposed framework was validated on cine bSSFP datasets acquired at 3T (uMR790, UIH, Shanghai, China). Quantitative cardiac functional analysis and blinded radiologist evaluation were performed to assess reconstruction accuracy and diagnostic quality.

Results: Compared with CRNN-based AI reconstruction for cardiac cine, the proposed framework produced cine images with sharper myocardial borders and reduced temporal blurring (acceleration ~15), as shown in Figure 2. Quantitative cardiac function analysis showed that ejection fraction, stroke volume, and cardiac output derived from the enhanced ACS images exhibited closer agreement with fully sampled references, as shown in Figure 3, agreement with radiologists’ quality ratings, assessed independently by three board-certified radiologists each with over 10 years of experience, further confirmed higher fidelity and greater diagnostic confidence.

Conclusion: Fast AI reconstructions suffer from temporal blurring and functional inaccuracy, and CS is limited by lengthy reconstruction times, whereas diffusion models^3-4 achieve high-quality results but remain slow and may introduce unrealistic details. By integrating AI with CS, our method preserves sharpness, improves functional accuracy, and enables fast, reliable cardiac cine MRI for clinical use.

Figure 1.
Figure 2.
Figure3