Prediction of Hepatic Parenchymal Change in Gd-EOB-DTPA MR Images after Stereotactic Body Radiation Therapy by Cycle GAN Deep Neural Network

Abstract

Stereotactic body radiation therapy (SBRT) is known to be an excellent ablative treatment option for small hepatocellular carcinoma (HCC), while possibly deteriorating hepatic function due to inevitable irradiation to surrounding non-tumorous liver parenchyma. A previous work attempted to reveal the threshold dose of parenchymal change after SBRT to be around 20 Gy by analyzing hypo-signal intensity on the hepatobiliary phase (HBP) of gadolinium ethoxybenzyl diethylenetriaminepentaacetic acid (Gd-EOB-DTPA)-enhanced magnetic resonance (MR) images. However, it found that estimating the parenchymal change is hard as the threshold dose can be patient-specific, ranging from 18 to 33 Gy. Thus, this work proposes to construct a prediction model for the hepatic parenchymal change after SBRT in Gd-EOB-DTPA enhanced MR images by using an image-based deep neural network. The deep learning with convolutional neural network encouraged numerous applications to the image-based prediction. This study aims to build an image-based deep-neural network that predicts the hepatic parenchymal change on HBP in follow-up enhanced MR images by correlating 3-channel input images of pre-treatment enhanced MR, CT and dose images to the output of the follow-up enhanced MR images. The recent advance in deep learning enabled for translating different types of medical imaging modalities by so-called generative adversarial network (GAN). The prediction accuracy could be further enhanced by placing additional generator in reverse direction to be robust to the image misalignment across different imaging modalities, which is named cycle-consistent GAN (cGAN). In this work, the proposed network based on cGAN was trained and tested with 36 and 7 patient cases, respectively. The prediction accuracy was assessed by comparing the predicted and enhanced follow-up MR images in root-mean-squared error (RMSE) and structural similarity (SSIM). For the seven testing cases, the image similarity was measured only inside the manually drawn liver contours, and 5 Gy iso-dose line. Inside the liver contours, the RMSE and SSIM between the reference follow-up MR and predicted MR images from cGAN were 591.21, and 0.9242, which outperformed the results from GAN. For the 5 Gy iso-dose lines, the structural similarity (SSIM) was reduced to 0.6511, which is mainly due to smoothed intensity transition in the predicted MR images. When visually checked, the locations with the signal intensity transition in predicted images correspond to the reference images. The image-based deep neural network with cGAN could possibly predict the hepatic parenchymal change in the follow-up enhanced MR images after SBRT for HCC.Abstract 2305; Table 1Comparison in prediction accuracy of GAN and cGAN deep neural networkLiver Contour>5Gy iso-dose lineRMSESSIMRMSESSIMGAN514.620.6217854.10.9127cGAN350.40.6511623.90.9242 Open table in a new tab