Click here to


Are you sure ?

Yes, do it No, cancel

Feasibility of Building a Deep Convolutional Neural Network for MR Image Based Synthetic-CT for Prostate Proton Planning: A Novel Method to Improve MRI Based Proton Dose Calculation Accuracy

X Ding*, S Chen , W Wang , A Qin , X Li , J Zhou , D Krauss , C Stevens , P Kabolizadeh , D Yan , William Beaumont Hospital, Royal Oak, MI


(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 2

Purpose: Magnetic Resonance Imaging (MRI) based proton planning faces bigger challenge compared to the photon planning due to the sensitivity of the range uncertainties. This study introduces a deep convolutional neural network for MRI based Synthetic-CT image (S-CT) generation and evaluated the dosimetry accuracy on prostate Intensity Modulated Proton Therapy (IMPT) planning.

Methods: A Paired CT and T2-weighted MR images were acquired from each of 36 prostate cancer patients. Ten CT-MRI pairs were used as tested set and the other 26 pairs as training set. The training subjects were augmented by applying 3-dimensional artificial deformations and then feed to a 2-dimensional U-net which contains 23 convolutional layers and 25.29 million trainable parameters. As a comparison to the U-net generated MRI synthetic CT images (S-CT_Unet), a multi-atlas deformable image registration based method was used to generate another set of S-CT images (S-CT_deform) for the ten tested CT-MRI pairs. Laterally opposed two-field IMPT plans with robustness optimization (3.5% and 5 mm) were created on the simulation CT images and then re-calculated on the S-CT_Unet and S-CT_deform respectively. Dose agreement was evaluated using 3D gamma criteria (3%/2 mm and 2%/2mm) and maximum point dose absolute difference within region of interests.

Results: Within the patient body, the 3%/2mm gamma passing rates were (96.71±2.92)% vs (96±2.88)%, p = 0.02, and the 2%/2mm gamma passing rates were (95.08±3.53)% vs (94.01±3.35)%, p = 0.008, with respect to the U-net vs the multi-atlas method. The corresponding maximum point dose absolute difference within CTV, femoral heads, rectum, and bladder were 6.4cGy vs 14.5cGy (p=0.31), 52.56cGy vs 77.0cGy (p=0.08), 72.9cGy vs 57.2cGy (p=0.32), and 15.9cGy vs 13.0cGy (p=0.29) respectively.

Conclusion: The U-net generated S-CT images significantly improved the gamma passing rates within the patient body compared to the multi-atlas deformable image registration method for prostate IMPT planning.


Not Applicable / None Entered.


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

Contact Email