Improving The Image Quality And Contouring Efficiency For Online Adaptive Radiotherapy Via Deep Learning

Objective: Precision radiotherapy can focus on the tumor target area and protect normal tissues,and online adaptive radiotherapy is an advanced technology to ensure its superior effect.However,there are still technical bottlenecks that limit its effective application: poor image quality of cone-beam computed tomography(CBCT)is not good enough for contouring and dose calculation;ineffectiveness of manual contouring affects treatment efficiency;CBCT cannot effectively monitor anatomical changes during the treatment;and the accuracy of dose distribution monitoring in proton adaptive radiotherapy need to be improved.In order to solve the above bottlenecks,this study introduced deep learning methods and improved the algorithms to effectively improve the efficiency of adaptive radiotherapy by combining them with clinical requirements.Methods: To address the problem of poor CBCT image quality,this study evaluates the advantages and disadvantages of existing supervised and unsupervised learning methods and proposes a two-step approach that combines their advantages: first,using modalities to obtain paired data for supervised learning training to remove image artifacts,and then using patient data for unsupervised learning training to improve the accuracy of CT values.To address the time-consuming and labor-intensive problem of manual contouring in adaptive radiotherapy,this study improved the deep learning method and established a personalized training strategy to effectively improve the contouring accuracy.Two sets of frameworks were designed for CBCT and Magnetic Resonance Imaging(MRI),respectively.To address the problem of poor image quality of intra-irradiation CBCT images acquired in rotational volume-modulated radiation therapy,this study proposed to model the scattering of radiation treatment sources to remove scattering and noise in the projection domain,and thus improved the quality of reconstructed images,so as to better monitor the changes of anatomical structures during patient treatment.To address the problem of insufficient accuracy of dose monitoring during proton adaptive radiotherapy,this study established a deep learning method to improve the accuracy of carbon and oxygen component calculation using dual-energy CT,thus improving the accuracy of the Positron emission tomography(PET)-based dose verificication.The effectiveness of this method was validated using Monte Carlo simulation.Results: The two-step method was effective in improving CBCT image quality to the CT level.It maintains the anatomical structure of CBCT compared to the supervised learning method alone.Our method improved the accuracy of CT values and effectively reduced artifacts compared to the unsupervised learning method alone,which was helpful for tumor identification.The personalized auto-conturing method significantly outperformed the conventional deep learning methods with following results of Dice similarity coefficients: for CBCT images of patients with nasopharyngeal carcinoma: 0.87 for primary tumor and 0.91 for clinical target area;for MRI images of patients with prostate cancer: 0.90 for clinical target area,0.96 for bladder,0.89 for rectum,and 0.94 for femoral head.The intra-irradiation CBCT image quality improvement method in the treatment reduced the mean absolute error from 86.3 to 49.8 HU.This study also tested tthe two scatter modeling techniques,and the results showed that both techniques were promising for improving image quality.The average absolute error of the results obtained by the predicted carbon-oxygen composition method is less than 2%,and this result was better than the conventional calculation methods,while having better robustness to noise.In the validation of Monte Carlo simulation,the PET signal activity distribution has an average relative error of 3-5%.Conclusion: In this study,improvements were made to address bottlenecks in adaptive radiotherapy,and the results obtained demonstrated the feasibility and effectiveness of these solutions.The effective application of deep learning methods to adaptive radiotherapy can improve the clinical process,increase the accuracy of radiotherapy and save treatment time.