Purpose: Patient-specific synthetic CT (sCT) generation based on MRI could allow for MRI-only radiation therapy, improving the clinical workflow. In this work, a deep-learning based method is developed for sCT generation based only on MRI.
Methods: We propose to incorporate cycle-consistent generative adversarial networks (cycle-GAN) for MRI-based synthetic CT estimation. A cycle-GAN is implemented to capture the relationship between CT and MRI images while also inversely supervising an MRI-to-CT transformation model. A dense-block architecture, which captures multi-scale information by extracting features from previous hidden layers and deeper hidden layers, is introduced to enhance algorithmic performance. Extracted patches of training MR images are fed into the generator (MRI-to-CT) to get the equal-sized sCT patches. Cylce-GAN relies on both generator and discriminator networks. The generator loss is computed by a combination of mean squared error between synthetic and real image, and between cycle and real image. The discriminator loss is computed by mean absolute error (MAE) between the synthetic and real images. Leave-one-out is used to evaluate the proposed algorithm, and generated sCTs are compared with original planning CT to quantitatively evaluate accuracy.
Results: This technique was validated with a clinical study of 20 brain patients. The MAE, peak signal-to-noise ratio (PSNR), and normalized cross-correlation (NCC) indexes were used to quantify estimation accuracy. Overall, the MAE, PSNR and NCC were 55.7 HU and 50.8 HU, 26.6 dB and 24.5 dB, and 0.96 and 0.93 on brain and pelvic site, respectively. Small MAE and high PSNR and NCC demonstrate the SCT prediction accuracy of the proposed learning-based method.
Conclusion: We have investigated a novel deep learning-based approach to accurately generate SCT from MRI. This CT image generation technique could be a useful tool for MRI-based radiation treatment planning or MRI-based PET attenuation correction of a PET/MRI scanner.
Funding Support, Disclosures, and Conflict of Interest: NIH R01 CA215718
Not Applicable / None Entered.