CBCT-to-CT Synthesis for Cervical Cancer Adaptive Radiotherapy via U-Net-Based Model Hierarchically Trained with Hybrid Dataset

Abstract

Purpose: To develop a deep learning framework based on a hybrid dataset to enhance the quality of CBCT images and obtain accurate HU values.

Materials and methods: A total of 228 cervical cancer patients treated in different LINACs were enrolled. We developed an encoder-decoder architecture with residual learning and skip connections. The model was hierarchically trained and validated on 5279 paired CBCT/planning CT images and tested on 1302 paired images. The mean absolute error (MAE), peak signal to noise ratio (PSNR), and structural similarity index (SSIM) were utilized to access the quality of the synthetic CT images generated by our model.

Results: The MAE between synthetic CT images generated by our model and planning CT was 10.93 HU, compared to 50.02 HU for the CBCT images. The PSNR increased from 27.79 dB to 33.91 dB, and the SSIM increased from 0.76 to 0.90. Compared with synthetic CT images generated by the convolution neural networks with residual blocks, our model had superior performance both in qualitative and quantitative aspects.

Conclusions: Our model could synthesize CT images with enhanced image quality and accurate HU values. The synthetic CT images preserved the edges of tissues well, which is important for downstream tasks in adaptive radiotherapy.

Keywords: adaptive radiotherapy; artifacts removal; cervical cancer; hierarchical training; image enhancement; synthetic CT.

Figure 1

Criteria of data selection. Initially, 180 and 50 patients were collected from Varian Trilogy and Elekta Axesse, respectively. After evaluation, 228 patients were enrolled in this study. One patient was excluded due to contrast agent usage when scanning the planning CT and another one was excluded due to failing in image pre-processing.

Figure 2

Flow chart of image pre-processing. The left are the original images and on the right are the processed images used to train the model. The window level/width was set to [−100, 400] HU. The whole image pre-processing included registration, removal of non-anatomical structures, resizing, cropping, adjusting, and translating.

Figure 3

Framework of deep learning model used in this study: (a) The overall architecture of the model. The model was developed on the basis of U-Net with residual blocks. The green box represents the convolution block, the blue box represents the residual block, the light orange box represents the up-sampling residual block, and the red box represents the convolution layer with filter size of 1 × 1. The number at the top of box is the filters in the layer. (b) Details of the convolution block, the residual block, and the up-sampling residual block shown in (a).

Figure 4

Illustration of the hierarchical training strategy. A total of three stages were included in this study. Except for the first stage, the other two stages were optimized on the basis of the previous stage. The multi-resolution image dataset for hierarchical training was generated by down-sampling the original CBCT images and pCT images by a factor of two.

Figure 5

Qualitative results of pseudo−CT images generated by different methods. Three randomly selected test cases are demonstrated in (A–C). The upper row shows the pCT, CBCT, and synthetic CT generated by our model and the RCNN model from the left to right. The bottom shows the difference maps for synthetic CT images when compared with pCT. The difference of HU values between synthetic CT generated by our model and pCT is minimal and the artifacts are obviously suppressed in all the three test cases.

Figure 5

Qualitative results of pseudo−CT images generated by different methods. Three randomly selected test cases are demonstrated in (A–C). The upper row shows the pCT, CBCT, and synthetic CT generated by our model and the RCNN model from the left to right. The bottom shows the difference maps for synthetic CT images when compared with pCT. The difference of HU values between synthetic CT generated by our model and pCT is minimal and the artifacts are obviously suppressed in all the three test cases.

Figure 6

HU line profiles and HU value histogram of test cases shown in Figure 5. For (A–C), the upper right is the HU line profile passing through the red dotted line on the left image for pCT, CBCT, and synthetic CT produced by our model and the RCNN model, respectively. The bottom left is the line profile passing through the yellow dotted line on the left image. The bottom right is the HU value histogram for the test case. The images generated by our model exhibit more similar trends in terms of HU value distribution corresponding to pCT when compared with the RCNN model and CBCT images.

Figure 6

HU line profiles and HU value histogram of test cases shown in Figure 5. For (A–C), the upper right is the HU line profile passing through the red dotted line on the left image for pCT, CBCT, and synthetic CT produced by our model and the RCNN model, respectively. The bottom left is the line profile passing through the yellow dotted line on the left image. The bottom right is the HU value histogram for the test case. The images generated by our model exhibit more similar trends in terms of HU value distribution corresponding to pCT when compared with the RCNN model and CBCT images.

Figure 7

Statistical analysis of quantitative results: (A) The comparison between different methods in terms of MAE. (B) Comparison of PSNR. (C) Comparison of SSIM. The synthetic CT images achieved improvement in all evaluation metrics compared with the original CBCT images. In addition, the established model outperformed the RCNN model.