Click here to


Are you sure ?

Yes, do it No, cancel

Evaluation of Synthetic CT Generation From CBCT Using a Deep Learning Model

A Haidari1,2*, D Granville2, E Ali1,2, (1) Carleton University, Ottawa, ON, CA, (2) The Ottawa Hospital Cancer Centre, Ottawa, ON, CA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Cone-beam computed tomography (CBCT) scans are used in image-guided radiation therapy for patient positioning and treatment adaptation. Due to lack of CT number accuracy (among other challenges), CBCT scans are not suitable for treatment planning purposes. Advances in deep learning have allowed for CT number recovery from CBCT via synthetic CT (sCT) generation. In this work, we perform an evaluation of a deep learning sCT generation model towards the purpose of conformal treatment planning based on CBCT-only scans.

Methods: For a cohort of 20 pelvic radiotherapy patients, sCT images were generated from CBCT scans using a deep learning model that was trained on independent data external to our centre. This model is part of the Elekta Advanced Medical Imaging Registration Engine (ADMIRE) software package, which is currently available only for research purposes. CBCT and sCT images were normalized and then registered to their respective treatment planning CT (TPCT) images that were used as the reference image (“ground truth”). For the CBCT volume within the external body contours for each patient, the CT number accuracy was quantified by taking the mean absolute error (MAE) for sCT-TPCT and CBCT-TPCT image pairs. Variations in patient positioning in the TPCT and CBCT were not accounted for.

Results: It was found that the MAE for sCT-TPCT was (39.5 ± 10.2) HU, and for CBCT-TPCT pairs was (224.3 ± 19.7) HU.

Conclusion: The ADMIRE sCT module shows promise for synthesizing TPCT-like quality images with accurate CT number reproduction. In future work, with a model that is trained on local data, we anticipate improved agreement between sCT and TPCT images.


Cone-beam CT, Dosimetry


IM- Cone Beam CT: Machine learning, computer vision

Contact Email