MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Joint Imaging-Therapy Scientific Session
  5. Deep Learning Approaches for Male Pelvic Synthetic CT Generation Using 2D and 3D Convolutional Neural Networks

Deep Learning Approaches for Male Pelvic Synthetic CT Generation Using 2D and 3D Convolutional Neural Networks

J Fu*, Y Yang , K Singhrao , D Ruan , A Kishan , C King , D Low , J Lewis , UCLA, Los Angeles, CA

Presentations

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

Room: Karl Dean Ballroom B1

Purpose: Accurate estimation of electron density maps and patient positioning reference images is an essential part of MR-only prostate radiotherapy. This can be achieved by generating synthetic CT (sCT) from MR images. Here, we developed 2D and 3D convolutional neural networks (CNNs) to generate male pelvic sCTs from individual T1-weighted MR images.

Methods: CTs and T1-weighted MR images of 20 prostate cancer patients were studied retrospectively. All CTs were deformably registered to MR images to generate paired 2D slices and 3D volumes for training the 2D CNN and the 3D CNN, respectively. The proposed 2D and 3D CNNs contained 27 convolutional layers. Residual shortcuts and instance normalization were implemented for fast convergence. A five-fold-cross-validation framework was used for generating all sCTs. CNN performance is evaluated by comparing generated sCTs to CTs using geometric and voxel-wise metrics.

Results: Each sCT generation took approximately 5.5 s. The mean absolute errors (MAEs) between sCTs and CTs were 40.5±5.4 HU and 37.6±5.1 HU for the 2D and 3D CNNs, respectively. The dice similarity coefficients (DSCs), recalls, and precisions for the bone region (thresholding the CT at 150 HU) were 0.81±0.04, 0.85±0.04, and 0.77±0.09 for the 2D CNN, and 0.82±0.04, 0.84±0.04, and 0.80±0.08 for the 3D CNN, respectively. P values of the Wilcoxon signed-rank tests for all metrics were less than 0.05 except for bone region recall (P=0.6).

Conclusion: The 2D and 3D CNNs generated accurate pelvic sCTs for the 20 patients. The generation speed was faster than other atlas-based and conventional machine learning methods. Statistical tests indicated that the 3D CNN was able to generate sCTs with better MAE, bone region DSC, and bone region precision. These results suggest that both CNNs have potential for enabling MR-only pelvic radiotherapy and MR-guided online adaptive planning. The sCT dosimetric accuary will be tested in the future.

Funding Support, Disclosures, and Conflict of Interest: This study was partially funded by the Varian master research agreement

Keywords

Computer Vision, MRI, Image-guided Therapy

Taxonomy

IM/TH- MRI in Radiation Therapy: Development (new technology and techniques)

Contact Email

Additional Links