257 - Deep Learning-Based Synthetic 3.0T High Field Magnetic Resonance Imaging for Brain Tumor Adaptive Radiotherapy in MR-Linac
Presenter(s)
B. Wang1, Y. Tang2, B. Liu3, J. Zhang1, Y. Liu4, B. Qiu3, X. J. Du3, H. Liu5, Y. Lu2, and X. W. Deng4; 1Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer,Sun Yat-sen University Cancer Center, Guangzhou, China, 2School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China, 3Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer,Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China, 4Department of Radiation Oncology, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer,Sun Yat-sen University Cancer Center, Guangzhou, China, 5State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China
Purpose/Objective(s): Magnetic resonance guided adaptive radiotherapy (MRgART) enables both inter-fractional and intra-fractional tracking of treatment target throughout the radiation treatment course. However, integrating magnetic resonance imaging (MRI) with linear accelerator system was challenging in engineering and often lead to compromise in static B0 field strength of the onboard MRI and in turn the quality of the online MRI images. To satisfy the need of the higher quality MRI images during online adaptations, we developed a hybrid cascade deep learning neuro-network to synthesize virtual 3.0T MRI images from the 1.5 T MR-Linac images for online adaptive radiotherapy (ART) of the brain.
Materials/Methods: Paired pretreat 1.5T MR-Linac images and 3.0T MR-sim images from 66 cases with brain metastasis were retrospectively collected for this study. Unpaired images included 368 online 1.5T MR-Linac image sets in the treatment course of 78 brain metastases cases who undergone MRI guided stereotactic radiotherapy (SRT) and 3.0T MR-sim image sets from 147 brain metastases cases who undergone stereotactic radiosurgery (SRS) or SRT treatment on conventional linear accelerators. A hybrid cascade GAN (HC-GAN) was established by combining conventional cyclic consistent GAN (cycleGAN) and conditional GAN (cGAN) networks. Synthetic 3.0T MRI was quantitively evaluated against real MR-sim images using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM). Clinical evaluation was also carried out by inviting 4 radiation oncologists to perform tumor delineation based on synthetic 3.0 T MRI (synT2w), MR-Linac images (T2w) and MR-sim images (T1C) of extra 18 brain metastasis cases respectively.
Results: HC-GAN significantly outperformed cycleGAN in terms of image quality of synthetic MRI. MAE dropped from 0.168 to 0.153 (P<0.001) for cycleGAN and HC-GAN respectively. PSNR increased from 24.43 dB to 25.56 dB (P<0.001) and SSIM increased from 0.832 to 0.859 (P<0.001) respectively. Delineation results has shown significant improvement of target delineation accuracy with synthetic 3.0T MRI compared to original 1.5T MR-Linac images. The mean dice similarity index (DSC) of GTVs contoured on synthetic 3.0T MRI and 1.5T MR-Linac images was 0.593 vs 0.526 compared to the baseline target respectively (P<0.001). For GTVs over 3 cc, the mean DSC reached 0.738 vs 0.695 for synthetic 3.0T MRI and 1.5T MR-Linac respectively.
Conclusion:
We proposed a novel HC-GAN network which could generate high quality 3.0T MRI images from 1.5T MR-Linac images. The structure, texture and detail of the image generated by this method was greatly improved compared with the traditional method. The addition of composite images could facilitate more accurate online target delineation for MRgART of brain metastasis.