Interventional cardiac MRI at 10 radial spokes per frame: KOSTAL reconstructs in under 34 ms at the quality of 7 to 11 times more sampling

Cardiac cine can now be reconstructed from only 10 radial spokes per frame at image quality close to much denser acquisitions, with under 34 ms of processing per frame, inside the 200 ms latency budget interventional cardiac MRI requires. Conventional gridding of the same 10 spokes reached a median SSIM of only 0.622, against 0.907 for KOSTAL; that quality corresponds to roughly 71 spokes of static gridding or 114 spokes of dynamic gridding. On 33 held-out test series from the open-access multi-coil OCMR dataset, SSIM and MSE were computed in a central 50% region of interest containing the heart against a 1000-spoke fully sampled radial reference, with paired two-sided Wilcoxon signed-rank tests and Holm correction.

Most of the gain comes from recurrence in k-space rather than image space: the k-space module alone reached a median SSIM of 0.891, while the same recurrent module placed directly in the image domain reached at most 0.650. Prior convolutional recurrent networks for video largely took image-domain inputs, and recurrence on k-space samples had not been reported; conventional view-sharing and keyhole methods merge new and old samples with fixed, hand-designed rules. Ablations compared plain convolutional RNN, convolutional LSTM and convolutional GRU cells in three configurations, k-space alone, image space alone and dual domain, with paired tests over the same 33 patients.

The recurrent state lets image quality recover quickly after an abrupt change of imaging plane, returning to steady-state quality within 75 ms by SSIM. Gridded reconstructions of comparable quality needed 111 ms at 75 spokes or 303 ms at 200 spokes. Thirty-three held-out long-axis-to-short-axis transition pairs were scored frame by frame against the 1000-spoke reference, with recovery time defined as the first time after the switch at which a reconstruction reached that pair's 10-spoke gridded steady state.

Without any retraining on a prospective cohort of 23 free-breathing pediatric patients, KOSTAL agreed best with the binned XD-GRASP reference, and its left-ventricular blood-pool volume curve lagged that reference by a median of 71 ms. Iterative reconstruction at 10 and 32 spokes lagged by 239 ms and 287 ms, and against clinical breath-held cine the end-diastolic volume bias was -0.7 mL. The prospective cohort covered a multi-slice short-axis stack in 23 patients and two-, three- and four-chamber long-axis views in 20, with volume agreement assessed over 69 short-axis slices from 23 patients by correlation, ICC and Bland-Altman bias.

The demonstrated scope is a held-out test on simulated golden-angle radial data from OCMR plus prospective free-breathing pediatric cardiac MRI acquisitions in 23 patients at a single center on a single vendor's scanners; reconstruction was performed offline, so latency was modeled from measured component times rather than observed in a live inline loop.

Training data were resampled from breath-held Cartesian bSSFP data with radial sampling simulated retrospectively; the prospective cohort lacks fully sampled ECG-gated golden-angle radial references and uses binned XD-GRASP instead; the validation cohort is small and from a single center and single vendor; arrhythmias were not directly evaluated; and whether the network would bias reconstruction of structurally atypical hearts toward a more normal appearance has not been tested directly.