PCA-Based Future Image Frame Prediction for Real-Time MRI-Guided Radiotherapy
Abstract
Purpose
MRI-radiotherapy hybrid systems enable real-time guidance of the therapeutic beam by dynamically adapting to tumor motion. However, inherent delays in imaging and beam control demand prediction of the tumor’s future position and shape for accurate targeting. We aim to predict intrafractional MR images 0.5–2.0 frames beyond the current image.
Methods
2D dynamic images from 10 lung and 10 liver patients were acquired using 3T MRI at 4 frames/s. A sliding window of 60 most recent frames was selected to characterize recent anatomical variations using principal component (PC) analysis. PC analysis was performed in k-space, and the 16 most significant PCs were used for prediction. The temporal trajectory formed by the PC scores was independently extrapolated using autoregression to 0.5–2.0 frames beyond the window. Using the 16 extrapolated scores as weights, the PCs were combined to generate predictive images. For verification, integer-predicted frames were compared against the acquired frames, and non-integer-predicted frames were compared against frames generated by an AI-based video frame interpolation method, whose image quality was evaluated using Natural Image Quality Evaluator (NIQE). The proposed method was benchmarked against an autoregressive model that predicts tumor centroid locations. For contour analysis, the same autocontouring technique was applied to both predicted and acquired images.
Results
The contour analysis yielded agreement with corresponding acquired images for 1- and 2-frame predictions, with average Dice coefficient of 0.94 and 0.90 for liver, and 0.90 and 0.88 for lung. Similar trend was found for the non-integer-predicted frames. The NIQE score for the AI-generated non-integer frames (5.93) agreed well with the acquired frames (5.41). The proposed method outperformed the autoregressive model, achieving 1–6% higher Dice coefficient. Average prediction time was 49 ms/frame.
Conclusion
The proposed method enables accurate and computationally efficient real-time image prediction to direct MLC motion ahead of anatomic motion in intrafractional MRIgRT.