| 研究実績の概要 |
X-ray beam parameter dependency of induced acoustic signals were reviewed analytically. Longer irradiation beam lengths (pulse duration >1μs) were shown to greatly attenuate magnitude of induced waves, complicating tomographic reconstruction. Hence, standard clinical radiotherapy system setups are not ideal with beam pulse lengths usually on order of ms.
3D acoustic wave simulations were performed. Significant artifacts were found proportional to distance from transducer elements as well to completeness in coverage of anatomy by distribution of sensor arrays. Attempts were made modeling US transducer designs as arcs and partial spheres more optimal for 3D imaging, but further exploration is necessary to find balance in a trade off between imaging error and sensors being physical obstructions themselves for the irradiating beam's path.
Focus was made on CT image reconstruction and artifact reduction for estimation of exact tissue properties real-time during therapy, a necessary component in parallel absolute dosimetric imaging. On-board and body surface depth imaging were used as motion surrogates to deform the higher quality pre-therapy planning CT. Several advanced algorithms to solve such inverse problems have been investigated, such as dictionary learning, neural representation learning, and Bayesian deep learning methods. Performance competitive, if not superior, to state of the art techniques were apparent in accuracy and inference speed. In the latter, reconstruction calculations took only fractions of a second. A significant advance toward realizing use in real-time.
|
