MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Conditional Generative Adversarial Networks (cGANs) for Magnetic Resonance (MR) Based Synthetic Computed Tomography (sCT) Images in Head and Neck Dose Calculations

P Klages*, J Jiang, R Farjam, JO Deasy, M Hunt, H Veeraraghavan, N Tyagi, Memorial Sloan Kettering Cancer Center, New York, NY

Presentations

(Thursday, 8/2/2018) 1:00 PM - 3:00 PM

Room: Karl Dean Ballroom B1

Purpose: To test the effectiveness of conditional generative adversarial (cGAN) based deep learning networks in generating synthetic CTs (sCTs) from MR images for dose planning in head and neck cancer patients.

Methods: Twelve head and neck cancer patients undergoing CT and MR simulation in radiotherapy treatment position were used to create a CT-MR atlas. All MR images in the atlas were preprocessed to remove intensity inhomogeneities and to standardize the intensity histograms. The Pix2Pix cGAN model was trained using (n=67634) pairs of CT-MR images consisting of image patches (128x128) from the training set of ten patients to generate sCT images from in-phase images of a dual echo 3D mDixon fast field echo MR images. Default parameter settings of Pix2Pix GAN were used and the training was stopped after 90 epochs. Preliminary algorithm accuracy was evaluated using the remaining two unused patient cases by (a) calculating Mean Absolute Error (MAE) between the sCT and original CT and (b) transferring structures and plan from planning CT to sCT and recalculating the dose.

Results: Preliminary results yield MAE of 120+/-200 HU and 130+/-200 HU for patients 1 and 2, respectively. A dosimetric comparison between the original planning CT and sCT plans showed agreement to within 1.5% for all structures. A qualitative evaluation between digitally reconstructed radiographs generated from sCT and original CT shows good accuracy, though as is common with MR to CT mappings, air and bone regions are misclassified in some regions. Synthetic image generation takes only 10s of seconds per data set using a trained network.

Conclusion: cGANs show promise for sCT generation for complex head and neck cancer anatomy. Additional testing on larger datasets and with more training data is required to see if the MR to CT mapping can be further improved at the air-tissue/bone-tissue interfaces.

Funding Support, Disclosures, and Conflict of Interest: This work was supported through a Research Agreement from Philips Healthcare

Keywords

MRI, CT, Dose

Taxonomy

IM/TH- MRI in Radiation Therapy: MRI for treatment planning

Contact Email