DiscoTSNet: a phase-guided temporal-spatial network for discontinuous respiratory motion prediction

Respiratory motion introduces positional uncertainties in radiotherapy for thoracic and abdominal tumors, and system latencies between respiratory signal acquisition and beam adjustment further increase targeting errors. This study aims to develop a precise real-time prediction method for discontinuous respiratory motion under prolonged latency conditions. Approach. We propose DiscoTSNet, a phase-guided temporal-spatial network for discontinuous respiratory motion prediction. The model combines a Phase-Guided Respiratory Decomposition module for separating respiratory signals into aperiodic, inhalation, and exhalation components; a Disco2D mechanism for organizing historical sequences according to known system latency; and a Spatial Feature Extraction module for modeling relationships among the coordinate channels of the surface markers. Main results. DiscoTSNet was evaluated on a publicly available CyberKnife respiratory motion dataset and achieved the lowest mean absolute error (MAE) and mean squared error (MSE) and the highest coefficient of determination (R^2) and correlation coefficient (CORR) among the compared models across the evaluated latency conditions from 200 ms to 1000 ms. At 1000 ms latency, its MAE, MSE, R^2, and CORR were 0.186, 0.107, 0.688, and 0.831, respectively. Significance. The proposed latency-guided discontinuous respiratory motion prediction method is tailored to radiotherapy motion compensation and has potential to improve real-time tumor tracking and enhance radiotherapy precision.