Supervised deep learning-based synthetic computed tomography from kilovoltage cone-beam computed tomography images for adaptive radiation therapy in head and neck cancer

Article information

Radiat Oncol J. 2024;42(3):181-191

Publication date (electronic) : 2024 May 30

doi : https://doi.org/10.3857/roj.2023.00584

Chirasak Khamfongkhruea^,1,2, Tipaporn Prakarnpilas¹, Sangutid Thongsawad^1,2, Aphisara Deeharing², Thananya Chanpanya², Thunpisit Mundee², Pattarakan Suwanbut², Kampheang Nimjaroen^1,2

¹Medical Physics Program, Princess Srisavangavadhana College of Medicine, Chulabhorn Royal Academy, Bangkok, Thailand

²Department of Radiation Oncology, Chulabhorn Hospital, Chulabhorn Royal Academy, Bangkok, Thailand

Correspondence: Chirasak Khamfongkhruea Medical Physics Program, Princess Srisavangavadhana College of Medicine, Chulabhorn Royal Academy, 906 Khamphaengpet 6 Rd., Lak Si, Bangkok 10210, Thailand. Tel: +66 2576 6038 E-mail: chirasak.kha@cra.ac.th

Received 2023 July 7; Revised 2024 February 24; Accepted 2024 February 29.

Abstract

Introduction

Head and neck cancer (HNC) has emerged as a prominent global health issue, ranking as the sixth most prevalent cancer globally. According to estimates from the Global Cancer Observatory, the overall incidence of HNC is on an ongoing upward trajectory, with predictions indicating an increase of 30% by the end of 2030 [1,2]. Due to the presence of multiple critical organs at risk (OARs) close to the treatment volume, radiation therapy commonly utilizes techniques such as intensity-modulated radiation therapy (IMRT) and volumetric-modulated arc therapy (VMAT) [3]. These advanced techniques enable precise targeting of the treatment volume while minimizing radiation exposure to surrounding critical structures. Nevertheless, the complexity of treatment techniques and anatomical changes can introduce uncertainty, leading to discrepancies between the planned and delivered doses. Thus, an innovative approach called adaptive RT (ART) has been proposed to address this challenge. ART involves adjusting the treatment plan to accommodate anatomical variations throughout the treatment course [4,5]. This is accomplished by assessing the patient's anatomical changes and re-calculating the dose distribution based on the updated images. Typically, these images can be obtained either through a repeat computed tomography (CT) scan or by utilizing images captured on the same day, such as kilovoltage cone-beam CT (kV-CBCT) [6]. Implementing the repeat CT scan technique would lead to an increase in the patient's radiation dose and require additional time and labor. Thus, utilizing kV-CBCT images for patient position alignment within the treatment room can effectively address these limitations. However, the direct application of kV-CBCT images for re-planning is constrained due to various factors, including inconsistent CT numbers, image quality issues related to scattered artifacts and noise, as well as limitations in the field-of-view [7-9]. Therefore, prior to utilizing kV-CBCT in ART, it is necessary to perform kV-CBCT correction.

In the existing literature, several techniques have been proposed for performing kV-CBCT dose calculation [8-10]. These methods encompass (1) the establishment of a calibration curve that correlates Hounsfield units (HU) with electron densities; this curve can be defined using either an adapted phantom or patient kV-CBCT images. However, it is essential to note that these methods are susceptible to image artifacts and patient scatter, which may impact their reliability [9]. (2) the utilization of density assignment methods, specifically bulk density override; this method involves segmenting an image into different tissue classes, such as soft tissues, air, and bones, and assigning appropriate densities to each class. When kV-CBCT images contain significant artifacts, considering alternative approaches can be a favorable option [11]. However, it relies on accurate structure segmentation and may result in an image with homogeneous tissues [9]. (3) the application of deformable image registration; the planning CT (pCT) images are deformed to kV-CBCT images, mainly including the deformation image registration and histogram matching methods. While this approach can effectively correct HU information in kV-CBCT images, it demands high accuracy in the image registration algorithm and matching method, especially in cases with substantial anatomical variations, such as tumor shrinkage or weight loss [12]. Additionally, this approach can be challenging due to inherent limitations in kV-CBCT imaging, including noise, low contrast, and a reduced field-of-view [13]. Finally, (4) the adoption of artificial intelligence algorithms to generate a synthetic CT (sCT); machine learning algorithms, specifically deep learning techniques, can be utilized to learn the mapping between kV-CBCT images and pCT images. These algorithms can learn intricate relationships and patterns through training on a vast dataset of paired kV-CBCT and pCT scans, enabling them to generate sCT images from kV-CBCT inputs [14]. Several studies have successfully demonstrated significant enhancements in CT number accuracy, image quality, and the mitigation of artifacts, thereby increasing the potential for the clinical implementation of ART [14-19]. Typically, there are two approaches for employing deep learning methods in sCT images. The first approach utilizes supervised training with paired images, often implementing algorithms like U-Net [20,21]. Conversely, the second approach involves unsupervised training with unpaired images, leveraging techniques such as generative adversarial networks (GANs) [14,16,22]. In order to integrate into clinical practice seamlessly, our research primarily concentrates on utilizing the supervised U-Net algorithm. This choice is motivated by its straightforward implementation, stable convergence, and rapid training process. Therefore, this study aims to generate and assess the performance of a U-Net deep learning-based algorithm in converting kV-CBCT scans to sCT images in terms of HU accuracy, image quality, and dosimetric accuracy similarity in the head and neck region.

Materials and Methods

3. Model architecture

The U-Net served as the basis for the proposed model in this study. The network model was developed using Keras and TensorFlow 2.9.0, utilizing Python version 3.8. The implementation involved NVIDIA CUDA Deep Neural Network Library version 8.1 and Compute Unified Device Architecture version 11.2. All experiments were performed on an NVIDIA Quadro RTX 8000 GPU with 48 GB of memory. The training process took place within a JetBrains PyCharm anaconda environment. The U-Net model, as illustrated in Fig. 1, follows a convolutional encoder–decoder network structure. The model utilizes pairs of kV-CBCT and pCT images as input and generates sCT images as output. The network architecture can be divided into two main phases: the encoding phase and the decoding phase.

Fig. 1.

U-Net network architecture: training (upper) and testing (lower). pCT, planning computed tomography; sCT, synthetic computed tomography; CBCT, cone-beam computed tomography; ReLU, rectified linear unit.

The model encodes by downscaling using 2D convolutions, rectified linear unit (ReLU) activation function, and maxpooling in six blocks: initial features of 32, increasing by 2; maxpooling of 2 × 2. Encoder reduces the image to 4 × 4 × 1024 with final 3 × 3 convolutions, 2048 features. Decoding uses transposed convolutions. Each up-sample block has a transposed convolution, concatenation, and two 3 × 3 convolutions; starting at 2048, features decrease by 2; decoder convolutions of 3 × 3, 32 features. The U-Net model is optimized using mean absolute error (MAE) loss.

Results

1. HU accuracy and image quality

Table 1 presents the MAE, RMSE, PSNR, and SSIM values of kV-CBCT and sCT images when compared to the reference pCT images. When comparing the metrics of the sCT to kV-CBCT images, there was an improvement in the values of MAE, RMSE, and PSNR. Specifically, the values decreased from 53.15 ± 40.09 to 41.47 ± 30.59 for MAE, from 153.99 ± 79.78 to 130.39 ± 78.06 for RMSE, and increased from 47.91 ± 4.98 dB to 49.93 ± 6.00 dB for PSNR, respectively. The SSIM of sCT images was 0.98 ± 0.02, which was higher compared to the SSIM of kV-CBCT images, which was 0.97 ± 0.02.

Table 1.

MAE, RMSE, PSNR, and SSIM values of kV-CBCT and sCT images

Fig. 2 showcases three axial slices of pCT, kV-CBCT, and sCT images. The sCT images demonstrate enhanced HU values that closely resemble those of pCT images while maintaining the geometric information present in kV-CBCT images. Furthermore, the sCT images effectively reduce streak artifacts observed in kV-CBCT images, as indicated by the red arrow in Fig. 2.

Fig. 2.

Comparison of three axial slices from planning CT, kV-CBCT, and synthetic CT images. In the synthetic CT images, indicated by the red arrow, a reduction in streak artifact is observed compared to the kV-CBCT images. CT, computed tomography; kV-CBCT, kilovoltage cone-beam computed tomography.

Fig. 3 illustrates the HU line profile across the body of pCT (orange), kV-CBCT (green), and sCT (blue) images. The line profiles indicate that the HU line of sCT images closely resembles that of pCT images, particularly at the boundaries of the body. sCT images exhibit reduced artifacts and smoother transitions compared to kV-CBCT images. Notably, at the interface between soft tissue and air, sCT images display a smoother edge compared to kV-CBCT images.

Fig. 3.

Axial images (upper) and Hounsfield unit line profiles (lower) of planning CT images (in orange), kV-CBCT images (in green), and synthetic CT images (in blue) acquired from independent datasets. CT, computed tomography; kV-CBCT, kilovoltage cone-beam computed tomography.

2. Dosimetric accuracy

Table 2 presents the 3D gamma passing rates concerning the dose distribution within sCT images, assessed using independent testing data. The evaluation was conducted utilizing gamma criteria of 2%/2 mm, 3%/2 mm, and 3%/3 mm. The outcomes reveal passing rates of 92.1% ± 3.8%, 93.8% ± 3.0%, and 96.9% ± 2.0%, respectively. These results highlight the remarkable performance of the 3D gamma passing rates in capturing the dose distribution within sCT images. The observed passing rate consistently maintains alignment with acceptable levels, even when subjected to slightly more stringent evaluation criteria.

Table 2.

The 3D gamma passing rates (%) of dose distribution in sCT images for independent testing data

In Fig. 4, box plots are depicted to showcase the percentage dose difference of target volume and OARs when comparing pCT and sCT images from eight independent patient datasets. Overall, the dose metric outcomes from the sCT-based plan closely resemble those obtained from the pCT-based plan. The percentage difference in dose for all structures experienced slight changes. The highest values fell within a 3% range, except for the parotid glands, where values exceeded 3% but remained below 5.2%. The observed deviation, which includes the presence of an anatomical change and the utilization of a bolus, can be attributed to patient #6, as depicted in Fig. 5.

Fig. 4.

The percentage dose difference between the planning CT and synthetic CT-based plans for the target and organs at risk (OARs). CT, computed tomography; PTV, planning target volume; Lt., left; Rt., right.

Fig. 5.

Comparison of dose distribution between the planning CT and synthetic CT-based plans for patients #6 (left) and #7 (right). The PTV 70 Gy is visually emphasized in red, while the PTV 59.4 Gy is represented in orange. The left and right parotid glands are depicted in blue and cyan, respectively. CT, computed tomography; PTV, planning target volume.

Fig. 5 presents the calculated dose distribution using pCT and sCT. In patient #7, the dose distribution exhibits a relatively similar pattern between the two image sets. However, in patient #6, the dose distribution differs significantly due to anatomical changes.

Fig. 6 showcases the DVH of structures comparing the pCT and sCT-based plans for patients #6 and #7. In patient #7, a slight dose difference was observed in the DVH. However, the dose difference between the pCT and sCT-based plans was more pronounced in patient #6. These findings highlight the impact of anatomical changes and the limitations of using kV-CBCT images for generating sCT, particularly in capturing accurate target volumes.

Fig. 6.

The comparison of dose-volume histogram (DVH) between the pCT and sCT-based plans for patients #6 (upper) and #7 (lower). Squared lines represent the DVH of the pCT-based plan, while the triangle lines represent the DVH of the sCT-based plan. pCT, planning computed tomography; sCT, synthetic computed tomography; PTV, planning target volume; Lt., left; Rt., right.

Discussion and Conclusion

This study aimed to investigate the model for generating sCT images from kV-CBCT scans using the supervised U-Net deep learning algorithm for ART treatment planning in the HNC region. This region is particularly susceptible to anatomical variations, such as weight and tumor size changes. The study focused on assessing the accuracy of HU, image quality, and dosimetric accuracy. Four evaluation metrics were employed to determine the performance of the generated sCT images. HU accuracy was evaluated using the MAE and RMSE, while image quality was assessed using PSNR and SSIM. In addition, the dosimetric accuracy was evaluated by measuring metrics such as D_95% and D_2% to assess the target volume and D_50% or D_max for the OARs.

Our study's findings are consistent with the results reported by Kida et al. [19], Li et al. [23], and Chen et al. [20]. In line with these studies, our results likewise indicated a reduction in MAE and RMSE values, accompanied by an enhancement in PSNR and SSIM values when compared to the original kV-CBCT images. Kida et al. [19] utilized a U-Net model structure similar to ours, excluding ReLU in transpose convolution, and employed MAE as the loss function in the pelvic regions. Our study demonstrates improved image quality regarding PSNR and SSIM, with values of 50.9 dB and 0.967, respectively. Comparing our study to the work of Li et al. [23], a 2D U-Net neural network in HNC was investigated. Our study surpasses their MAE, PSNR, and SSIM performance. However, RMSE was not considered in their work. They reported MAE, PSNR, and SSIM values of 56.89 ± 13.84, 28.80 ± 2.46 dB, and 0.71 ± 0.032, respectively. However, it is essential to note that Chen et al. [20] achieved better HU accuracy metrics results than our study by employing a loss function that combined structure dissimilarity and MAE. They reported an average MAE, RMSE, PSNR, and SSIM of 18.98, 60.16, 33.26 dB, and 0.8911, respectively. Among the observed results, the most notable finding is that our study achieved the highest SSIM value. This indicates higher similarity and quality between our generated sCT images and the reference CT images (pCT). The results depicted in Figs. 2 and 3 demonstrated that the sCT images achieved better image quality. Notably, the presence of scatter artifacts was effectively eliminated, leading to images free from disturbances. Furthermore, the smoothness of the synthesized images was found to be comparable to that of pCT images.

According to the HU line profiles of pCT, kV-CBCT, and sCT images, the sCT images not only successfully eliminated scatter artifacts but also demonstrated a higher degree of similarity in HU values to the pCT images compared to the kV-CBCT images, especially at the body boundaries. At the thyroid cartilage region, all sCT images showed better intensity with the reduction from the peak of kV-CBCT images. The smoothness of soft tissue had improved in all sCT images. Although when the HU line crosses through the surrounding edge of the air region, such as the outside of the body and subglottic larynx, it indicates that sCT images had rounded corners close to pCT images.

In our analysis of dosimetry, we found that our study had a lower gamma passing rate of 2%, 2 mm criteria compared to another study. Specifically, our results indicated this rate to be 92.1% ± 3.8%. Jihong et al. [24] reported a rate of 95.7% ± 1.9% for sCT images with uncorrected CBCT and 97.1% ± 1.9% for sCT images with corrected CBCT. Meanwhile, Yoo et al. [25] achieved an impressive result of 99.7% ± 0.0% for sCT using a combination loss functions for model training. Both studies utilized advanced deep learning techniques, with Jihong et al. [24] employing unsupervised learning via CycleGAN with HU correction, and Yoo et al. [25] enhancing performance through the integration of perceptual loss into L1 and structural similarity loss functions during model training. These findings suggest that unsupervised deep learning and specialized loss functions can enhance the quality of sCT images, and preprocessing techniques such as HU correction can further improve outcomes.

It is important to note that the presence of registration errors in the training data can lead to inaccuracies in the model's performance as it may be trained to generate erroneous image predictions. In addition, the anatomical changes between the pCT and sCT images may influence the results of the dose evaluation. Further analysis revealed that the outlier data points were primarily associated with one specific patient, precisely patient #6. This particular patient exhibited significant anatomical changes. The parotid gland of patient #6 showed higher dose difference values compared to the other patients. The observed dose uncertainties can be attributed to the fact that the anatomical changes in the target volume of patients.

The U-Net model presents numerous benefits, such as reducing global scattering and enhancing local HU accuracy by integrating global and local features within the image's spatial domain [20]. Moreover, this model offers advantages such as straightforward implementation, stable convergence, and the fastest training [26]. Nevertheless, a limitation of the model lies in determining the optimal number of depths in the encoder–decoder network, which depends on the complexity of the training task. Additionally, the design of skip connections between the encoder and decoder networks lacks a robust theoretical framework.

While our study demonstrated improvements in HU accuracy, image quality, and dosimetric accuracy, several limitations should be acknowledged. Firstly, the limited field-of-view of kV-CBCT images restricted the availability of anatomical information, particularly for low-risk target volumes in the shoulder region of HNC. This limitation could affect the accuracy of the generated sCT images in those areas. Secondly, our study had a relatively small image data sample size for training and testing. A more extensive and diverse dataset could yield even better results and provide a more robust evaluation of the model's performance accuracy. Additionally, a strategic approach to dataset enhancement through image augmentation techniques involving translation, inversion, and slight rotation presents a viable avenue for further improvement. Thirdly, anatomical changes may occur between the pCT and kV-CBCT images due to the time gap between image acquisitions (usually taken on separate days). These anatomical variations could potentially impact the training of the U-Net algorithm, which relies on paired image data. Therefore, exploring unsupervised models such as GANs could enhance the model's performance, particularly when dealing with unpaired image data (intra-individual co-registration) [18,24-30]. Furthermore, the process of rigidly registering the CT and CBCT images might not have been adequate in establishing the required image similarity for network training. This concern could be mitigated by incorporating deformable image registration, which could have been essential for enhancing image similarity to a greater extent. In addition, the model developed in this study is based on a 2D approach, generating sCT images slice-by-slice. To further enhance the performance of the 2D model, prior research [20] has proposed the utilization of a volumetric neural network operating in three dimensions. This could involve extracting image features more precisely and effectively. Despite these limitations, the 2D supervised U-Net model remains a valuable tool for improving the accuracy and quality of the sCT generation from kV-CBCT images, effectively addressing the challenges posed by anatomical variations in HNC patients.

In conclusion, this study used supervised U-Net deep learning to generate sCT images from kV-CBCT scans in patients with HNC. Our evaluation focused on assessing the HU accuracy, image quality, and dosimetric accuracy of the sCT images. The sCT images produced by the U-Net model successfully reduced artifacts and noise while preserving the anatomical structure obtained from the kV-CBCT images. The HU accuracy and image quality of the sCT images were close to those of the pCT images. Furthermore, regarding dosimetric accuracy, the structures in the sCT images, including PTV, and OARs, closely resembled their counterparts in the pCT images. However, it is essential to acknowledge that anatomical changes can occur between acquiring pCT and kV-CBCT images. Our findings suggest that the U-Net model holds significant potential in generating kV-CBCT-based sCT images for ART in HNC patients. This approach offers a convenient method of obtaining updated anatomical information without subjecting the patient to additional radiation doses.

Article information Continued