Adaptive recursive Bayesian estimation using expectation maximization for respiratory motion correction in Nuclear Medicine

A method to correct for irregular, non stationary respiratory motion is required to improve quantitative and qualitative accuracy of Nuclear Medicine Images. Solutions to date rely on temporally regular respiratory motion with static models learnt from training data. An adaptive approach with dynamic parameter learning of motion models is required. To this avail we cast respiratory motion estimation as a Hidden Markov model. An expectation maximization based Kalman smoother algorithm is utilized to infer hidden states of motion from observations of the patient´s chest motion alone. The framework is validated using a computational anthropomorphic phantom (XCAT) with seven respiratory cycles with varying amplitude and frequency. A PET study is simulated with four 16mm lung lesions to assess the effectiveness of the approach. Preliminary tests are also performed on dynamic MRI data of a single volunteer. The likelihood of dynamical model fitting is monitored for individual respiratory cycles. Optimal estimates of previously unseen motion are made using the Kalman smoother. The proposed method can correct for respiratory motion to the order of 1.5mm. A thirty percent increase in mean uptake value for the corrected tumors in the simulated PET study was observed.