Dataset and image preprocessing

The preliminary data processing in this study was based on the analysis of the original data, which included 1477 head and neck CT slices and 1477 CBCT slices obtained from a cohort of 20 different patients. The original CBCT images were processed using the Varian iCBCT technology for noise reduction during image acquisition. The tube voltage for CT was 120 kVp and CBCT was 100 kVp. The voxel size of the CT images was 0.511 mm × 0.511 mm × 1.989 mm with an image resolution of 512×512. Further, the voxel size of CBCT images was 0.976 mm × 0.976 mm × 3 mm with an image resolution of 512 × 512.

The acquisition time of CBCT images is about one month after that of CT images. Owing to the different acquisition times of the CT and CBCT images, there was a minor difference in deformation. Therefore, image registration between CT and CBCT images is required. The anatomical deformation registration technique was selected in 3D Slicer software and optimized using an adaptive stochastic gradient descent algorithm [39]. The CT images were registered to the CBCT images to maintain the anatomical geometry consistent, and the registered CT images were resampled to a resolution of 0.976 mm × 0.976 mm × 3 mm, the same with the CBCT images.

In this study, we strictly followed the registration results to produce CT-CBCT image pairs for training and testing. We randomly selected 1267 CT-CBCT image pairs from 17 patients as the training datasets and 210 CBCT-CT image pairs from the remaining three patients as the testing datasets where each patient dataset contained approximately 70 pairs.

Network structure

Figure 1 shows the network structure of CBAM-U-Net. The input data for the CBAM-U-Net network were CBCT images, and the output data were SCT images that were processed by the neural network. The first layer of the network was used to expand the original CBCT data to 32 channels. Using the down-sampling module, the height and width of the data were reduced to half of the input image, and the number of channels was doubled. After 6-layer down-sampling modules, 8 × 8 × 1024 feature images were created. Contrary to the function of the downsampling module, the feature images were processed by the upsampling module, which doubled the feature image height and width and reduced the number of channels by half. The 8 × 8 × 1024 feature images were up-sampled using 6-layer up-sampling modules, processing the feature image dimensions to 512 × 512 × 32. Finally, an SCT image with dimensions 512 × 512 × 1 was produced by the output module. The attention mechanism modules were integrated into both the downsampling and upsampling modules. These modules enabled the network to allocate calculation weights to the sampling modules in each layer, thereby enhancing the accuracy of the neural network outputs.

Fig. 1Full-screen View

(Color online) Overall structure of the CBAM-U-Net network

The weight calculation of the feature image was executed in the sampling module of each layer to increase the output accuracy of the neural network. The module of the convolutional attention mechanism is illustrated in Fig. 2. The convolutional block attention module comprised two submodules: the channel and spatial attention modules. Each feature image was assigned a weight parameter using a module called the channel attention mechanism module. The spatial attention mechanism module then processed the feature image by applying weight parameters to each pixel of the feature image. This process occurred during feature-image generation. To increase the accuracy of the HU values in the output results, the convolutional block attention module provided additional weight parameters for certain feature images and areas within all feature images.

Fig. 2Full-screen View

Convolutional block attention module structure

A description of the channel attention module is presented in Fig. 3. The input feature images were initially processed using an adaptive average pool, with the output having dimensions of 1 × 1 × C, where C is the total number of channels. The weight parameters were generated after the input data passed through two linear layers and a Sigmoid layer. The input image was multiplied by the weight parameters to produce the output at the end of the module for the channel-attention mechanism. During the training of a neural network, certain feature images significantly affected the quality of the ideal output SCT image. After passing through the channel-attention mechanism module, these significant feature images were assigned larger weights.

Fig. 3Full-screen View

Channel attention mechanism module structure

Figure 4 shows the spatial attention mechanism module. Subsequent to the channel attention mechanism module, the spatial attention mechanism module accepted the output of the channel attention mechanism module as its input. The spatial attention mechanism module compressed the feature images into a one-channel image and processed the feature image using an average pooling layer and a maximum pooling layer. The image maps of the average and maximum pooling layers were concatenated and processed using a convolution layer. The feature images were multiplied by the spatial weight to obtain the final output of the spatial attention mechanism module. Therefore, the spatial attention mechanism module can assign higher weights to key areas of the feature images and lower weights to areas outside the key areas to improve the capability of the network. The channel and spatial attention mechanism modules both assigned calculation weights to the feature images of the network. Throughout the learning process, the calculation weights were continuously modified to generate improved SCT images.

Fig. 4Full-screen View

Spatial attention mechanism module structure

Network training parameter

PyTorch was used as the computing framework, and L1loss (Eq. 1) was chosen as the loss function. The model training engines were two NVIDIA RTX 2080Ti with 24 GB of computational memory. The batch size was set to 8, and each GPU received 4 images as the input. The initial learning rate was set at 1×10^-4, and there were 180 training epochs. Adam was used as the optimizer and StepLR was used as the learning rate adaptor. Training was performed on both the sCTU-Net and CBAM-U-Net models with the same training parameters, and the training weight parameters were saved for each network. $L1Loss\left(S\left(x_i\right)\mathrm{,}{y}_i\right)=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left|S\left(x_i\right)-y_i\right|}$ (1) where S(xi) is the pixel value of the generated SCT image, yi is the pixel value of the corresponding CT image, and n is the total number of pixels in the image.

Image evaluation

To evaluate the noise reduction performance, the MAE, RMSE, PSNR, and SSIM were used. The ground truth images were registered as CT images. The values of the data from the three testing datasets were analyzed using these four evaluation parameters, and the average values of the three testing datasets were calculated as the final assessment parameter.

The MAE and RMSE are expressed in Eqs. 2 and 3, respectively. Both are the most prevalent assessment criteria used in image processing and offer the advantages of being easily understood and quantifiable. The smaller the MAE and RMSE of the two images, the higher the similarity between them. The corresponding formulas are as follows: $MAE\left(S\left(x_i\right)\mathrm{,}{y}_i\right)=\frac{1}{m}{\displaystyle \sum _{i=1}^m\left|S\left(x_i\right)-y_i\right|}$ (2) $RMSE\left(S\left(x_i\right)\mathrm{,}{y}_i\right)=\sqrt{\frac{1}{m}{\displaystyle \sum _{i=1}^m{\left(S\left(x_i\right)-y_i\right)}^2}}$ (3) where S(xi) is the pixel value of the generated SCT image, yi is the pixel value of the corresponding CT image, and m denotes the total number of image pixels.

PSNR is defined as the ratio between the maximum possible value of a signal and distorting noise power that affects the quality of its representation. The larger the PSNR, the higher is the similarity between the two images. The formula for PSNR is expressed as Eq. 4: $PSNR=10{{\displaystyle \log }}_{10}\left(\frac{Ma{x}_i^2}{MSE}\right)$ (4) where Maxi is the maximum value of the image pixel sampling point and MSE is the mean square error.

SSIM is a statistic used to determine the overall similarity between two images. It is commonly used as a statistic to measure image generation and processing. The dynamic range of SSIM ranges from -1 to 1. The closer the SSIM is to 1, the higher is the similarity in the total information content between the two images. SSIM was calculated using Eq. 5: $SSIM\left(x\mathrm{,}y\right)=\frac{\left(2{\mu }_x{\mu }_y+C_1\right)\left(2{\sigma }_{xy}+C_2\right)}{\left({\mu }_x^2+{\mu }_y^2+C_1\right)\left({\sigma }_x^2+{\sigma }_y^2+C_2\right)}$ (5) where μx is the mean of x, μy is the mean value of y, ${\sigma }_x$ is the variance of x, ${\sigma }_y$ is the variance of y, ${\sigma }_{xy}$ is the covariance of x and y, C₁=(k₁L)², C₂=(k₂L)² is the constant used to maintain stability, and L is the maximum value of image pixels. Further, k₁ and k₂ were 0.01 and 0.03, respectively.

To show the matching degree between SCT, CBCT, and CT images in different HU value ranges. Q-Q (quantile-quantile) plots [40] between SCT and CT images and between CBCT and CT images from samples of the testing data were also plotted for detailed evaluation. Q-Q plots are commonly used in mathematical statistics to test the consistency of the distribution of two groups of data. The closer the two sets of data were distributed, the closer the Q-Q plot of the two groups of data was to 45 from the reference line.

Evaluation of dose calculation accuracy

The matRad dose calculation engine, an open-source software developed by the German Cancer Research Center (DKFZ, Heidelberg) for scientific purposes of radiotherapy planning, was used to calculate the dose distribution of proton beams, where the ray-tracing algorithm was used for dose calculation [41].

First, for the depth dose analysis, the IDD curve analysis of a proton beam with a nominal energy of 114.5 MeV is used in matRad. The dose grid was set as 1 mm × 1 mm × 1.5 mm. A proton beam in air with a full width at half maximum (FWHM) of 5 mm was applied to the dose calculation model of the head and neck CT, CBCT, CBAM-SCT, and U-Net-SCT images of the patient testing datasets. Second, for the lateral dose profile analysis, CBCT, CBAM-SCT, and U-Net-SCT images were analyzed using a square-field γ-index pass rate to compare the calculation results. The scanned field size was set to 4 cm × 4 cm, the spot spacing was 2 mm, and the square-field dose calculation employed proton beams with nominal energy of 114.5 MeV. Third, for further lateral dose analysis, a simple plan was also implemented, and under the same radiation conditions, the dose distribution of CT was used as a reference to compare their γ-index pass rates.

Comparative analysis of CT images

The network weight parameters of CBAM-U-Net and sCTU-Net were obtained under the same training conditions. The weight parameters produced by the sCTU-Net and CBAM-U-Net models were utilized to generate the SCT images of the three patient datasets. The data obtained from the generation of SCT images were evaluated using the four indicators mentioned above. Figure 5 shows representative transversal images of CT, CBCT, and SCT from a patient. Both the sCTU-Net and CBAM-U-Net networks effectively reduced image noise and restored the original CT image features in regions with CBCT image artifacts. The quantitative analysis mean value data are presented in Table 1.

Fig. 5Full-screen View

Typical transverse planes of (a) CT, (b) CBCT, (c) U-Net-SCT, and (d) CBAM-SCT images of patient

Table 1 demonstrates that the U-Net-SCT and CBAM-SCT images were significantly superior in the four evaluation metrics compared to the CBCT images. Moreover, CBAM-SCT images exhibited better image quality than the U-Net-SCT images.

To demonstrate the CBAM-U-Net’s ability to reduce noise, error distribution maps were obtained by subtracting the CT images from CBCT and SCT images, as shown on the left side of Fig. 6. The CBCT, and CBAM-SCT images had an MAE values of 23.68, 17.75, and 13.02 HU, respectively. The image errors generated by CBCT were worse than those generated by SCT, which mainly appeared around the bone border. CBCT images also contained severe scattered noise. CBAM-U-Net and original sCTU-Net both reduced scatter artifacts and controlled noise to an acceptable level. Moreover, CBAM-U-Net exhibited a better noise reduction ability than sCTU-Net, particularly in high-density areas such as the skull.

Fig. 6Full-screen View

(Color online) Absolute HU value difference of between CT and CBCT, SCT images. (left column). Q-Q plot between CT and CBCT images, and between CT and U-Net-SCT images, and between CT and CBAM-SCT images. (right column)

Q-Q plots between CBCT and CT images and between SCT and CT images from one test patient are shown in Fig. 6. The range of the horizontal coordinates of the Q-Q plots was set as the HU value range of the CT image, whereas that of the vertical coordinates of the Q-Q plot was set as the HU value range of CBCT or SCT. The Q-Q plot of CT-CBCT deviated from the 45 reference line, particularly in high-density tissue. Whereas, sCTU-Net corrected most of the errors, although there were still errors in high-density tissue areas. The Q-Q plot of CT-CBAM along the reference line of 45, regardless of the soft tissue and bone areas, demonstrated that CBAM-SCT images had the same data distribution as the CT images. The Q-Q plots show that the CBAM-U-Net network exhibited a better noise reduction ability in high-density tissue areas than the original sCTU-Net network.

Single-beam depth dose analysis

The IDD curve of the proton beam was analyzed using matRad. The high-density and soft-tissue areas of the three tested patients were selected as the analysis areas. The selection of these two areas reflected the actual clinical treatment conditions. Moreover, in dose calculations, the artifacts of high-density tissue areas are commonly the cause of large dose calculation errors. Figure 7 shows the dose distribution map for Patient 1 along the beam is shown in Fig. 7.

Fig. 7Full-screen View

(Color online) (a) CT images high density areas dose distribution of patient 1; (b) CBCT images high density areas dose distribution of patient 1; (c) U-Net-SCT images high density areas dose distribution of patient 1; (d) CBAM-SCT images high density areas dose distribution of patient 1; (e) CT images soft tissue areas dose distribution of patient 1; (f) CBCT images soft tissue areas dose distribution of patient 1; (g) U-Net-SCT images soft tissue areas dose distribution of patient 1; and (h) CBAM-SCT images soft tissue areas dose distribution of patient 1;

The corresponding IDD curves are presented in Fig. 8. The IDD curves of the CT images were used as the reference curves. The IDD curves for the high-density and soft-tissue areas were compared.

Fig. 8Full-screen View

(Color online) (a) Complete IDD curve in high density areas; (b)partial IDD curve in high density areas; (c) complete IDD curve in soft tissue areas; and (d)partial IDD curve in soft tissue areas

To show more details of the IDD curve at the Bragg peak region, for the IDD curve of the high-density areas, the depth 100 mm–145 mm IDD curve was selected for analysis, and for the IDD curve of the soft tissue areas, the depth 120 mm–145 mm IDD curve was selected for analysis. Both sCTU-Net and CBAM-U-Net exhibited effective noise reduction in soft tissue areas, as shown in Fig. 8, and their IDD profiles were comparable. However, the noise reduction effect of CBAM-U-Net was better than that of the original sCTU-Net in high-density areas. The CBAM-U-Net IDD curve was closer to the CT’s IDD curve than to that of sCTU-Net. To quantify this difference, Table 2 presents the depth at the peak of IDD curves for the three patients.

Because of the image artifacts in CBCT images, the IDD curves of the CBCT exhibited considerable inaccuracy compared to those of CT. Moreover, the results of the IDD curves of CBCT cannot reflect the anatomical structure information of patients from the perspective of proton radiotherapy. The IDD curves of CBAM-SCT and U-Net-SCT exhibited a better degree of matching than that of CBCT. Whereas, those of CBAM-SCT and U-Net-SCT had the same profile as the IDD curves of CT. According to the IDD curves, the CBAM-U-Net network was more effective than the original sCTU-Net network in reducing noise in high-density tissues. In addition, the results of CBAM-SCT images with a single-energy beam from the treatment angle and other key angles reflected tissue changes at the radiation site and the degree of anatomical deformation of the related tissue.

Lateral dose comparison analysis

In matRad, the dose results of CT were used as a reference. CBCT, CBAM-SCT, and U-Net-SCT images were analyzed using the 1%/1mm and 3%/3mm γ-index criteria in a square field to compare the calculation results. The square field size was set to 4 cm × 4 cm, and the spot spacing was 2 mm. The experiment employed proton beams with nominal energy of 114.5 MeV. The high-density boundary was selected as the radiation analysis area because it can assess the difference in dose distribution between the high-density and soft tissue areas. Figure 9 shows the scanned field lateral dose distribution and absolute dose difference of Patient 1 compared with the dose distribution in the CT images. The γ-index pass rates of CBCT, CBAM-SCT, U-Net-SCT, and CT for the three patients are listed in Table 3.

Fig. 9Full-screen View

(Color online) (a) Dose distribution in CT image square fields of patient 1; (b)dose distribution in CBCT images square field of patient 1; (c)dose distribution in U-Net-SCT image square fields of patient 1; (d)dose distribution in CBAM-SCT images square field of patient 1; (e)absolute dose difference between CBCT and CT image square fields of patient 1; (f)absolute dose difference between U-Net-SCT and CT image square fields of patient 1; and (g)absolute dose difference between CBAM-SCT and CT image square fields of patient 1

A smaller absolute dose difference existed between the CBAM-SCT and CT image square fields than that between the CBCT, U-Net-SCT, and CT images. In addition, Figure 9 shows that the absolute dose difference of CBAM-SCT images in the bone region was lower. The γ-index pass rate of CBAM-SCT was superior to that of the original sCTU-Net network under the criteria of 1%/1 mm and 3%/3 mm γ-index pass rates. In particular, under the 1%/1 mm calculation condition for patient 1, the γ-index pass rate of CBAM-SCT was 82.23%, which is considerably higher than that of U-Net-SCT’s 79.03%. Noise in higher-density tissues with a higher stopping power ratio leads to a greater dose error in radiotherapy dose calculation. CBAM-U-Net has a more robust noise reduction function in high-density areas; therefore, CBAM-SCT images are more suitable for clinical dose calculation requirements than the original sCTU-Net network.

According to the results of the square-field calculation, the CBAM-U-Net network can improve the accuracy of the dose calculation and eliminate the image artifacts existing in the original CBCT images. The γ-index pass rate of the SCT images generated by the CBAM-U-Net network was higher than that produced by the original sCTU-Net network. The application of square-field dose calculations can partially consider the influence of anatomical structural changes on the original treatment plan and provide additional data support for subsequent treatment.

To further evaluate the effect of image quality on the dose calculation results, delineation of a simple PTV and OAR was implemented. CT, CBCT, U-Net-SCT, and CBAM-SCT used the same PTV and OAR contours as in Patient 1 to eliminate errors caused by manual delineation. A single-field plan was adopted with a 180° gantry angle and 0° couch angle, and single-field inverse planning was used to optimize the dose distribution. The squared deviation function was used for dose optimization in the PTV area and the squared overdosing function was used for dose optimization in the OAR. The results of the CT images were selected as references, and the CBAM-SCT, and U-Net-SCT images were analyzed using the 2%/2mm γ-index criterion to compare the dose distribution. The absolute dose differences are shown in Fig. 10.

Fig. 10Full-screen View

(Color online) (a) Dose distribution in CT images of patient 1; (b)dose distribution in CBCT images of patient 1; (c)dose distribution in U-Net-SCT images of patient 1; (d)dose distribution in CBAM-SCT images of patient 1; (e)absolute dose difference between CBCT and CT images of patient 1; (f)absolute dose difference between U-Net-SCT and CT images of patient 1; and (g)absolute dose difference between CBAM-SCT and CT images of patient 1;

The γ-index pass rate of CBAM-SCT was 88.97%, which is better than that of U-Net-SCT and CBCT (88.06% and 77.01%, respectively). Using the CBAM-U-Net network to eliminate CBCT image artifacts may provide a foundation for proton-adaptive therapy.