Click here to


Are you sure ?

Yes, do it No, cancel

Inter-Vendor Compatibility and Transfer Learning for MR-Based Synthetic CT Deep Learning Models for Domain Adaptation

P Klages*, N Tyagi, H Veeraraghavan, Memorial Sloan-Kettering Cancer Center, New York, NY


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

Room: AAPM ePoster Library

Purpose: To evaluate the applicability of a trained synthetic CT model on similar sequence images from another vendor, and to investigate transfer-learning techniques as a method to improve sCT generation to both sequence sets when initial model training sets are unavailable.

Methods: For initial training, 30 cranial SRS radiotherapy patients with Philips T1w SPGR post contrast MRIs and diagnostic CT images taken within 24 hours were used. The images were rigidly aligned and used to train a 2.5D patch-based pix2pix neural network using 6-fold cross-validation; training was performed for 100 epochs over 1-week. For testing, two additional sets were used: 14 pairs of Philips SPGR Post Sim MR and CT images, and 10 pairs of GE BRAVO MR and CT images. All images were standardized to the same reference MR image from the training set using ITK. Transfer learning was applied for the GE test set, and a mix of Philips and GE data to observe how transformation accuracy changed with training epoch.

Results: The mean absolute errors (MAEs) for the cross-validated training set trained to 100 epochs was 103.7+/-14.6 HU; MAEs were (110.1+/-12.7 HU, 123.9+/-9.2 HU) for the Philips and GE test sets, respectively. With transfer learning on the GE BRAVO set, sCT artifacts were visibly reduced and MAE decreased to 113.8+/-14.3 HU, though the Philips test set MAE increased to 131.3+/-13.2 HU. Using transfer learning with a mix of the test sets, MAE was reduced for all test cases (106.5+/-10.4 HU, 112.7+/-11.7 HU for Philips, GE, respectively). Convergence using transfer-learning based training occurred within hours (by 10 epochs) for our training parameters.

Conclusion: Vendor-specific MR image differences are distinct enough to cause visible artifacts for models trained to a specific vendor, but these results suggest transfer learning may refine parameter weights for both modalities with limited supplemental data.

Funding Support, Disclosures, and Conflict of Interest: This study was supported under a Master Research Agreement with Philips Health Care


MRI, Transformation, Image Processing


IM/TH- MRI in Radiation Therapy: Synthetic CT

Contact Email