Optimal hyperparameters of CBCT-based synthetic CT using U-net deep learning to improve image quality for adaptive radiotherapy in the H&N region
Keywords:
synthetic CT, cone-beam CT, deep learning, U-net, hyperparametersAbstract
Background: Cone-beam CT (CBCT) imaging is used for adaptive radiation therapy (ART) in head and neck cancer (HNC) due to its more convenient image acquisition and no additional dose. However, CBCT limitations in Hounsfield (HU) accuracy and image quality have emerged for treatment planning. Recently, several studies have proposed using deep learning to generate synthetic CT (sCT) images from CBCT images. However, the quality of images depends on the hyperparameter setting.
Objectives: To determine the optimal hyperparameters of the U-net deep learning (DL) for generating sCT images for ART in HNC.
Material and methods: To generate sCT images, U-net DL with a mean absolute error loss function was used in this study. A total of 3491 image pairs from pCT and CBCT datasets from 40 HNC patients were split into 80% (2976 images from 32 patients) and 20% (515 images from 8 patients) for training and testing, respectively. Each parameter for tuning the U-net model, consisting of learning rates, batch sizes, and epochs, was investigated with various hyperparameter settings in a total of 45 conditions. The best model was assessed using four metrics, including a mean absolute error (MAE) and root mean square error (RMSE) for HU accuracy, peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) for image quality between sCT and pCT images, as well as a training time.
Results: For optimal hyperparameters, we found that the learning rate was set to 1e-3, batch size of 8, and epoch of 200. According to this setting, MAE, RMSE, and PSNR improved from 53.15 ± 40.09, 153.99 ± 79.78, and 47.91 ± 4.98 to 41.47 ± 30.59, 130.39 ± 78.06, and 49.93 ± 6.00, respectively, while SSIM remained constant. The learning rate played an essential role in the training model. All models with various hyperparameters enhanced the reduction of artifacts and noise. The edges of the bone and the soft tissue boundary were clearly visible. The average training time of an optimal hyperparameter was 6 hours and 36.6 minutes (398 ms/step), while it took less than 10 seconds to generate sCT images.
Conclusion: Hyperparameter optimization can improve the quality of sCT images for treatment planning. This study demonstrates the potential of U-net to use CBCT images for ART in HNC.
References
Spencer S. Head and neck cancers: Advantages of advanced radiation therapy and importance of supportive care. Journal of the National Comprehensive Cancer Network. 2018 May 1;16(5S):666-9. doi: 10.6004/jnccn.2018.0049.
Belshaw L, Agnew CE, Irvine DM, Rooney KP, McGarry CK. Adaptive radiotherapy for head and neck cancer reduces the requirement for rescans during treatment due to spinal cord dose. Radiation Oncology. 2019 Nov 1;14(1): 1-7. doi: 10.1186/s13014-019-1400-3.
Yeh SA. Radiotherapy for head and neck cancer. Semin in Plastic Surgery. 2010 May 21;24(02):127–136. doi: 10.1055/s-0030-1255330.
Castadot P, Lee JA, Geets X, Grégoire V. adaptive radiotherapy of head and neck cancer. Vol. 20, Seminars in Radiation Oncology. 2010 Apr 1;20(02):84–93. doi: 10.1016/j.semradonc.2009.11.002.
Schwartz DL. Current progress in adaptive radiation therapy for head and neck cancer. Curr Oncol Rep. 2012 Apr;14(2):139–47. doi: 10.1007/s11912-012-0221-4.
Sonke JJ, Aznar M, Rasch C. Adaptive radiotherapy for anatomical changes. Semin Radiat Oncol. 2019;29(3):245–57. doi: 10.1016/j.semradonc.2019.02.007.
Posiewnik M, Piotrowski T. A review of cone-beam CT applications for adaptive radiotherapy of prostate cancer. Physica Medica. 2019;59:13–21. doi: 10.1016/j.ejmp.2019.02.014.
Søvik Å, Rødal J, Skogmo HK, Lervåg C, Eilertsen K, Malinen E. Adaptive radiotherapy based on contrast enhanced cone beam CT imaging. Acta Oncol (Madr). 2010 Oct 1;49(7):972–7. doi: 10.3109/0284186X.2010.498433.
Buchanan A, Mcdavid D. Characterization and correction of cupping effect artefacts in cone beam CT. Dentomaxillofac Radiol. 2012 Mar 1;41:217–23. doi: 10.1259/dmfr/19015946.
Jaju P, Jain M, Singh A, Gupta A. Artefacts in cone beam CT. Open J Stomatol. 2013 Jan 1;03:292–7. 2013;03(05):292–7. doi: 10.4236/ojst.2013.35049.
Schulze R, Heil U, Groß D, Bruellmann D, Dranischnikow E, Schwanecke U, et al. Artefacts in CBCT: a review. Dentomaxillofac Radiol. 2011 Jul 1;40:265–73. doi: 10.1259/dmfr/30642039.
Kida S, Nakamoto T, Nakano M, Nawa K, Haga A, Kotoku J, et al. Cone beam computed tomography image quality improvement using a deep convolutional neural network. Cureus. 2018 Apr 29;10(4). doi: 10.7759/cureus.2548.
Chen L, Liang X, Shen C, Jiang S, Wang J. Synthetic CT generation from CBCT images via deep learning. Med Phys. 2020 Mar 1;47(3):1115–25. doi: 10.1002/mp.13978.
Barateau A, de Crevoisier R, Largent A, Mylona E, Perichon N, Castelli J, et al. Comparison of CBCT-based dose calculation methods in head and neck cancer radiotherapy: from Hounsfield unit to density calibration curve to deep learning. Med Phys. 2020 Oct 1;47(10):4683–93. doi: 10.1002/mp.14387.
Vincent P, Larochelle H, Lajoie I, Bengio Y, Manzagol PA, Bottou L. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research. 2010 Dec 1;11(12). doi: 10.5555/1756006.1953039.
Han X. MR-based synthetic CT generation using a deep convolutional neural network method. Med Phys. 2017 Feb 1;44. doi: 10.1002/mp.12155.
Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. International Conference on Machine Learning. 2015 Jun 1;37:448-456. doi: 10.5555/3045118.3045167.
Nair V, Hinton GE. Rectified linear units improve restricted boltzmann machines. International Conference on Machine Learning. 2010 Jan 1;807–14. doi: 10.5555/3104322.3104425.
Smith LN. A disciplined approach to neural network hyper-parameters: Part 1 -- learning rate, batch size, momentum, and weight decay. arXiv preprint arXiv:1803.09820. 2018 Mar 26. doi: 10.48550/arXiv.1803.09820.
Rashmi, Ghose U, Gupta M. Comparative design analysis of optimized learning rate for convolutional neural network. In: Sharma H, Saraswat M, Kumar S, Bansal JC, editors. Intelligent Learning for Computer Vision. Singapore: Springer Singapore; 2021;339–52. doi: 10.1007/978-981-33-4582-9_26.
Qian X, Klabjan D. The impact of the mini-batch size on the variance of gradients in stochastic gradient descent. arXiv preprint arXiv:2004.13146. 2020 Apr 27. doi: 10.48550/arXiv.2004.13146.
Afaq S, Rao S. Significance of epochs on training a neural network. International Journal of Scientific & Technology Research. 2020 Jun;9(06):485-8.
Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. International Conference on Medical image computing and computer-assisted intervention. 2015 Oct 5;234-241. doi: 10.1007/978-3-319-24574-4_28.
Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. Unet++: redesigning skip connections to exploit multiscale features in image segmentation. IEEE Trans Med Imaging. 2020 Jun;39(6):1856–67. doi: 10.1109/TMI.2019.2959609.
Krogh A, Hertz JA. A simple weight decay can improve generalization. International Conference on Neural Information Processing Systems. 1991;4:950–7. doi: 10.5555/2986916.2987033.
Bergstra J, Bengio Y. Random search for hyper-parameter optimization. Journal of machine learning research. 2012 Feb 1;13(2). doi: 10.5555/2188385.2188395.
Li Y, Zhu J, Liu Z, Teng J, Xie Q, Zhang L, Liu X, Shi J, Chen L. A preliminary study of using a deep convolution neural network to generate synthesized CT images based on CBCT for adaptive radiotherapy of nasopharyngeal carcinoma. Physics in Medicine & Biology. 2019 Jul 16;64(14):145010. doi: 10.1088/1361-6560/ab2770.
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2023 Thai Association of Radiation Oncology
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
บทความที่ได้รับการตีพิมพ์เป็นลิขสิทธิ์ของวารสารมะเร็งวิวัฒน์ ข้อความที่ปรากฏในบทความแต่ละเรื่องในวารสารวิชาการเล่มนี้เป็นความคิดเห็นส่วนตัวของผู้เขียนแต่ละท่านไม่เกี่ยวข้องกับ และบุคคลากรท่านอื่น ๆ ใน สมาคมฯ แต่อย่างใด ความรับผิดชอบองค์ประกอบทั้งหมดของบทความแต่ละเรื่องเป็นของผู้เขียนแต่ละท่าน หากมีความผิดพลาดใดๆ ผู้เขียนแต่ละท่านจะรับผิดชอบบทความของตนเองแต่ผู้เดียว
