Room: Karl Dean Ballroom B1
Purpose: We propose a deep-learning based method for pseudo-CT image synthesis from its corresponding MR image.
Methods: We used CT and T1-weighted MR image pairs of nineteen glioblastoma patients from the Cancer Imaging Archive (TCIA) as experimental data. The CT-MR image pairs were aligned by rigid-body registration using SimpleITK. The data oversampling was done by augmentation (flip and rotation) to overcome small amount of data. Finally, the dataset was divided into three categories: train (60%), validation (20%), and test sets (20%). The Conditional GAN model was trained by collecting 2D MR-CT slice pairs from each subject's 3D MR/CT volumetric data. The model consists of two parts: G (generator) is a fully connected network to generate pseudo CT images G(x, z) from a random noise vector z under the condition x (corresponding MR image), and D (discriminator) is a convolutional neural network to discriminate between an real CT image (ground truth) and an estimated pseudo-CT image. The G tries to minimize objective function against the D which tries to maximize it. Thus, the model can be trained to learn a pix-to-pix mapping to predict pseudo CT images from their corresponding MR images. The proposed model was implemented with Pytorch. The geometric accuracy of the pseudo-CT images generated from the test MR image set was evaluated against the real CT image using dice similarity score.
Results: Generating a pseudo-CT slice for each test MR image slice with our trained model only took <1 s while the training process took 5 hours. The overall average dice similarity score was 0.934 for all test cases.
Conclusion: This result showed that our method is feasible for predicting pseudo-CT images from their corresponding MR images and is applicable for MR-only treatment planning as well.