MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Image Synthesis and Automation of Synthetic CT From MRI Data for MRI-Driven RT

S Mousavi1*, A Hosni-Abdalaty2, D Shultz3, A Berlin4, C Coolens5, T Stanescu6, (1) Princess Margaret Cancer Centre, ,,(2) Princess Margaret Cancer Centre, Toronto, ,CA, (3) Princess Margaret Cancer Centre, ,,(4) Princess Margaret Cancer Centre, ,,(5) Princess Margaret Hospital, Toronto, ON, (6) Princess Margaret Cancer Centre, Toronto, ON, CA

Presentations

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

Room: AAPM ePoster Library

Purpose: Development, optimization, validation and practical implementation of deep learning models for synthetic CT (SynCT) applied to radiotherapy.

Methods: T1-T2-weighted MR sets were used to investigate two Generative Adversarial Networks (GAN) - cGAN and CycleGAN. 50/50 training and 10/10 validation sets from CNS/H&N cases were processed. CT and MR voxel resolution was 1 mm³. CNS data was acquired using mask on MR and CT. H&N data had 3/5-point mask on for CT and no mask for MR (no RF-coil clearance), resulting in a significant inter-procedure neck displacement. Multiple parameters were tested to optimize the networks/models: a) resolution (128-256-512), b) data formatting (transverse-sagittal-coronal), c) sample size (25-50 sets), d) contrast (T1/T2/mixed). For CycleGAN, training was done with paired/unpaired sets. CT-to-SynCT sets were compared by a) similarity metrics MAE/PSNR, b) RT plan comparison and c) 3D-boney surface rendering (registration/conformality). For clinical deployment, GAN models were implemented on an automated imaging pipeline (QUANTMR) for live SynCT generation. QUANTMR also enforced image quality validation of MR data using a predictive model.

Results: Our results showed that a good SynCT model can be achieved after training on only 25 cases. Training up to 400 epochs showed that approximately 200 epochs were sufficient to achieve a robust solution and avoid overfitting. PSNR/MAE values were 34/0.004 and 31/0.006 for cGAN and CycleGAN, respectively, when paired sets were used for training. Unpaired training did not significantly improved the model performance for CNS, but showed a superior performance in H&N. Metrics showed a modest improvement when training was increased to include 50 cases.

Conclusion: GANs for SynCT need to be carefully chosen to match the data available for training. Reliable models can be built with tens of CT-MR sets given adequate data quality and preparation. The study successfully developed and implemented a GAN-based infrastructure for SynCT in CNS-H&N RT.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email