Non‐stationary transformers‐based model for predicting liver motion for interleaved two‐dimensional cine magnetic resonance imaging

Abstract Background Magnetic resonance (MR)‐guided online adaptive radiation therapy (MRgOART) has attracted increasing attention owing to its capabilities to create daily adaptation plans and perform real‐time motion management. However, during multileaf collimator (MLC) tracking, MRgOART systems face a latency of ∼300 ms, which can compromise treatment accuracy. Therefore, predictive models for respiratory‐induced liver motion are essential to compensate for system delays and enable precise MLC tracking. Purpose This study aims to develop a respiratory motion prediction model based on non‐stationary transformers (NsTransformers) using cine MR images acquired using interleaved imaging on the Elekta Unity system. The performance of the proposed model is compared with that of iTransformers, bidirectional long short‐term memory (LSTM) with an attention mechanism (biLSTM‐ATT), LSTM and linear regression models to evaluate its potential for real‐time clinical applications. Methods Seventeen liver cancer patients treated with MRgOART were enrolled. Coronal and sagittal cine MR images were selected from three‐plane interleaved images during free breathing or abdominal compression. Respiratory motion was defined as the displacement of the centroid position of the liver obtained via intensity‐based deformable image registration. Data augmentation techniques, including random noise addition, amplitude modulation, and frequency transformation, were used to expand the training dataset. The NsTransformers model was trained to predict future centroid positions at forecasting horizons of 200, 400, and 600 ms. The root mean square error (RMSE) and margin‐based accuracy were used as evaluation metrics, and the statistical significance of the models was assessed using the Friedman and Nemenyi tests. Results The NsTransformers model consistently outperformed all comparative models across all forecast intervals. The RMSE results of the NsTransformers model were superior to those of the other models, with the NsTransformers model demonstrating statistically significant improvements ( p < 0.05) compared to the other models. In addition, the NsTransformers model achieved a higher margin‐based accuracy across multiple prediction margins. The computation time of the NsTransformers model was ∼5 ms per prediction, which is sufficiently short for real‐time applications. However, the prediction accuracy degraded under conditions of irregular respiratory motion. Conclusion A motion prediction model based on NsTransformers that effectively predicts respiratory‐induced liver motion from interleaved cine MR images was developed. The model demonstrated superior prediction accuracy compared with all comparative models and holds promise in compensating for latency in MRgOART MLC tracking.