Validation of deep learning-based CT image reconstruction for treatment planning

Deep learning-based CT image reconstruction (DLR) is a state-of-the-art method for obtaining CT images. This study aimed to evaluate the usefulness of DLR in radiotherapy. Data were acquired using a large-bore CT system and an electron density phantom for radiotherapy. We compared the CT values, image noise, and CT value-to-electron density conversion table of DLR and hybrid iterative reconstruction (H-IR) for various doses. Further, we evaluated three DLR reconstruction strength patterns (Mild, Standard, and Strong). The variations of CT values of DLR and H-IR were large at low doses, and the difference in average CT values was insignificant with less than 10 HU at doses of 100 mAs and above. DLR showed less change in CT values and smaller image noise relative to H-IR. The noise-reduction effect was particularly large in the low-dose region. The difference in image noise between DLR Mild and Standard/Strong was large, suggesting the usefulness of reconstruction intensities higher than Mild. DLR showed stable CT values and low image noise for various materials, even at low doses; particularly for Standard or Strong, the reduction in image noise was significant. These findings indicate the usefulness of DLR in treatment planning using large-bore CT systems.


Discussion
This study aimed to assess the usefulness of DLR (AiCE) in treatment planning using large-bore CT.The current study found that AiCE showed less change in CT values and smaller SDs than AIDR.These results suggest that AiCE can be used in radiation therapy as well as AIDR, which has been widely used, and it can be expected to play a role in recommending uncertainty in dose calculations because of the low noise in the images.The noise  reduction effect was particularly large in the low-dose region (Fig. 3b), indicating that AiCE may contribute to dose reduction in CT for treatment planning.Previous studies have shown that DLR algorithms reduce noise [4][5][6][7][8][9] , and the results of this study follow similarly.
It is well known that there are variations in organ delineation, and high-quality images can reduce delineation uncertainty in treatment planning [22][23][24][25] .This is a phantom-based study, despite demonstrating little change in CT values; furthermore, combined with the benefit of improved image quality, DLR may be useful for delineation.The AiCE Mild reconstruction showed little improvement in SD, suggesting that the reconstruction intensity should be stronger than that of Mild for treatment planning CT.The reconstruction kernel of AiCE in this study only considered the Body_sharp filter, and a multifaceted study is required.Images with low variability are also important for particle therapy.In particle therapy, CT images are used for range estimation 26,27 .DLR has been shown to reduce SD, and this method can be applied to particle therapy by evaluating the accuracy of calculating the stopping power.
Dose reduction in treatment-planning CT is another important topic [28][29][30] .Remarkably, in this study, AiCE showed stable performance, even at low doses.The difference between AiCE and AIDR became larger for lowdensity materials (Fig. 2b), but this was caused by the fluctuation of the CT values of AIDR (Figs. 2 and 4b).The CT values of AiCE showed very little change with dose (Fig. 4), and at low doses, an SD reduction was observed at all reconstruction intensities (Fig. 3).The difference in SD between Mild and Standard/Strong was large, again suggesting the usefulness of reconstruction intensities that were higher than Standard.
The principal limitation of this study is that we only validated the CT-ED table at the pre-treatment planning stage and did not examine the clinical impact of differences in CT reconstruction images.Furthermore, we did not clarify the contribution of the DLR to the quality of treatment planning and dose reduction in CT treatment planning.However, the usefulness of DLR in reducing noise in clinical and low-dose CT images has been verified.This study showed a small variation in CT values and small SD in pre-treatment planning using large-bore CT; based on these results, there may be little change in dose distribution in treatment planning using patient CT data, which will contribute to the reduction of exposure dose and contouring variation owing to improved image quality.More research using patient CT data to use the DLR for treatment planning is required.

Conclusions
This study is the first to demonstrate that the AiCE DLR algorithm is useful in radiotherapy with large-bore CT through the comparison with AIDR.Using a CT-ED tablet phantom, we found that AiCE showed stable CT values and low SD for various materials even at low doses; particularly for Standard or Strong, the reduction in SD was significant.The AiCE deep neural network trained on patient data can be applied to radiation therapy by confirming the findings obtained with the simple geometric phantom using anthropomorphic phantoms and patient datasets.Notwithstanding these limitations, this study suggests that DLR can be useful for treatment planning using large-bore CT systems.

Materials and methods
CT system and principles of an image reconstruction algorithm.Data were acquired using an Aquilion Exceed LB (Canon Medical).This CT has a 90 cm wide bore and wide field of view (FOV) and is available for image reconstruction using DLR and H-IR.AiCE uses artificial intelligence nodes called "neurons", which are networked in multiple layers to mimic human neuron connections.A deep convolutional neural network (DCNN), consisting of layers of neurons, is trained to perform complex tasks and generate CT images.The DCNN input is analyzed by several network layers, referred to as "hidden layers".The hidden layers contain "convolutional layers", in which the neurons act as feature selectors on small patches of data.The DCNN in AiCE has thousands of neurons and samples the feature space.For the successful performance of the DCNN, the network structure must be optimized, which affects the image quality and reconstruction speed.To achieve the best computational efficiency and improve image quality, network structure elements such as the number of network layers, the number of neurons in each layer, and convolution kernel size have been optimized in AiCE.The engine was trained using a sample dataset containing both high-and low-quality input data compared to gold standard images, allowing the system to distinguish signal from noise and reconstruct superior signal-to-noise ratios with the same radiation dose as conventional scanning techniques.To output the best optimal results, millions of image pairs were used in the training of AiCE DLR, and they were validated by thousands of phantom and patient images.
The H-IR algorithm used in this study is Adaptive Iterative Dose Reduction (AIDR).AIDR is a threedimensional CT algorithm that reduces image noise while preserving image quality 20,21 .AIDR is a commercial H-IR algorithm that combines reconstruction and denoising in the raw data and image space domains, AIDR manipulates the projection data using noise and scanner statistical models to adjust for variations in the patient, scan parameters, and scanner itself, and adapts the filter intensity based on the relative noise levels.

Data acquisition and reconstruction parameters.
The tube voltage was 120 kV, and the tube currents were 265, 200, 100, and 25 mA.The detector configuration was 0.5 mm × 80-rows (160-slices), the rotation time was 0.5 s/rotation, and the field-of-view was LL size (500 mm).The reconstruction kernel and reconstruction strength of the AIDR were the same as those used in clinical settings (reconstruction kernel: FC13, reconstruction strength: Mild).The reconstruction kernel for AiCE was body-sharp, and the reconstruction strength was compared with three parameters: Mild, Standard, and Strong.The scan parameters are presented in Table 1.

Phantom and data analysis. A large-diameter CT-ED table phantom (Advanced Electron Density
Phantom; SUN NUCLEAR) was used to measure the CT values and image noise of the various materials.The rods of the advanced electron density phantom mimic water, cortical/inner bone, lung, and liver.They are highly equivalent to the medical standards for human tissue densities and can optimally convert CT values to electron density.Figure 5 shows an overview and analytical view of the phantom.In this study, 10 materials were used, ranging from the lung (electron density: 0.288 c/cm 3 ) to the cortical bone (electron density: 1.774 c/cm 3 ).Moreover, regions of interest (ROI) of mimicked water (electron density: 0.999 c/cm 3 ) were created at the top, bottom, left, and right sides for analysis.The ROI size used in the analysis was approximately 300 mm 2 , and the ROI was set at the center of each material.CT values and variations (SD) within the ROI were analyzed.The analysis was performed using the viewer provided with the CT device.The CT value is the mean of the CT values within the specified ROI, The SD is the analysis of the variations (1σ) of CT value within the ROI.

Figure 1 .
Figure 1.Difference in CT values between AIDR and AiCE at various doses.The difference in CT values was obtained by the difference between AIDR and AiCE; the results for the three reconstruction intensities of AiCE are shown.

Figure 2 .
Figure 2. Difference in CT values between AIDR and AiCE for each material at 265 mAs (a) and 25 mAs (b).The difference in CT values was obtained by the difference between AIDR and AiCE; the results for the three reconstruction intensities of AiCE are shown.

Figure 3 .
Figure 3. Relative SD of AiCE to AIDR for 265 mAs (a) and 25 mAs (b) for each material.Where Difference was calculated as AiCE-AIDR.

Figure 4 .
Figure 4. CT value-to-electron density conversion table at 265 mAs (a) and 25 mAs (b).AIDR and AiCE of the three reconstruction intensities, all four data are shown.

Table 1 .
Summary of scan parameter.