Generating missing patient anatomy from partially acquired cone-beam computed tomography images using deep learning: a proof of concept.

Autor: Shields B; Biomedical Technology Services, Townsville University Hospital, Townsville, Australia. benjamin.shields@health.qld.gov.au.; School of Chemistry and Physics, Queensland University of Technology, Brisbane, Australia. benjamin.shields@health.qld.gov.au., Ramachandran P; School of Chemistry and Physics, Queensland University of Technology, Brisbane, Australia.; Department of Radiation Oncology, Princess Alexandra Hospital, Brisbane, Australia.
Jazyk: angličtina
Zdroj: Physical and engineering sciences in medicine [Phys Eng Sci Med] 2023 Sep; Vol. 46 (3), pp. 1321-1330. Date of Electronic Publication: 2023 Jul 18.
DOI: 10.1007/s13246-023-01302-y
Abstrakt: The patient setup technique currently in practice in most radiotherapy departments utilises on-couch cone-beam computed tomography (CBCT) imaging. Patients are positioned on the treatment couch using visual markers, followed by fine adjustments to the treatment couch position depending on the shift observed between the computed tomography (CT) image acquired for treatment planning and the CBCT image acquired immediately before commencing treatment. The field of view of CBCT images is limited to the size of the kV imager which leads to the acquisition of partial CBCT scans for lateralised tumors. The cone-beam geometry results in high amounts of streaking artifacts and in conjunction with limited anatomical information reduces the registration accuracy between planning CT and the CBCT image. This study proposes a methodology that can improve radiotherapy patient setup CBCT images by removing streaking artifacts and generating the missing patient anatomy with patient-specific precision. This research was split into two separate studies. In Study A, synthetic CBCT (sCBCT) data was created and used to train two machine learning models, one for removing streaking artifacts and the other for generating the missing patient anatomy. In Study B, planning CT and on-couch CBCT data from several patients was used to train a base model, from which a transfer of learning was performed using imagery from a single patient, producing a patient-specific model. The models developed for Study A performed well at removing streaking artifacts and generating the missing anatomy. The outputs yielded in Study B show that the model understands the individual patient and can generate the missing anatomy from partial CBCT datasets. The outputs generated demonstrate that there is utility in the proposed methodology which could improve the patient setup and ultimately lead to improving overall treatment quality.
(© 2023. Crown.)
Databáze: MEDLINE
Externí odkaz: