Prediction Of Response In Radiation Therapy Analytical Models And Modelling

Author : Lise Wei
Publisher :
ISBN 13 :
Total Pages : 145 pages
Book Rating : 4.6/5 (846 download)

DOWNLOAD NOW!

Download or read book Medical Image Analytics (radiomics) with Machine/deeping Learning for Outcome Modeling in Radiation Oncology written by Lise Wei and published by . This book was released on 2020 with total page 145 pages. Available in PDF, EPUB and Kindle. Book excerpt: Image-based quantitative analysis (radiomics) has gained great attention recently. Radiomics possesses promising potentials to be applied in the clinical practice of radiotherapy and to provide personalized healthcare for cancer patients. However, there are several challenges along the way that this thesis will attempt to address. Specifically, this thesis focuses on the investigation of repeatability and reproducibility of radiomics features, the development of new machine/deep learning models, and combining these for robust outcomes modeling and their applications in radiotherapy. Radiomics features suffer from robustness issues when applied to outcome modeling problems, especially in head and neck computed tomography (CT) images. These images tend to contain streak artifacts due to patients’ dental implants. To investigate the influence of artifacts for radiomics modeling performance, we firstly developed an automatic artifact detection algorithm using gradient-based hand-crafted features. Then, comparing the radiomics models trained on ‘clean’ and ‘contaminated’ datasets. The second project focused on using hand-crafted radiomics features and conventional machine learning methods for the prediction of overall response and progression-free survival for Y90 treated liver cancer patients. By identifying robust features and embedding prior knowledge in the engineered radiomics features and using bootstrapped LASSO to select robust features, we trained imaging and dose based models for the desired clinical endpoints, highlighting the complementary nature of this information in Y90 outcomes prediction. Combining hand-crafted and machine learnt features can take advantage of both expert domain knowledge and advanced data-driven approaches (e.g., deep learning). Thus, we proposed a new variational autoencoder network framework that modeled radiomics features, clinical factors, and raw CT images for the prediction of intrahepatic recurrence-free and overall survival for hepatocellular carcinoma (HCC) patients in this third project. The proposed approach was compared with widely used Cox proportional hazard model for survival analysis. Our proposed methods achieved significant improvement in terms of the prediction using the c-index metric highlighting the value of advanced modeling techniques in learning from limited and heterogeneous information in actuarial prediction of outcomes. Advances in stereotactic radiation therapy (SBRT) has led to excellent local tumor control with limited toxicities for HCC patients, but intrahepatic recurrence still remains prevalent. As an extension of the third project, we not only hope to predict the time to intrahepatic recurrence, but also the location where the tumor might recur. This will be clinically beneficial for better intervention and optimizing decision making during the process of radiotherapy treatment planning. To address this challenging task, firstly, we proposed an unsupervised registration neural network to register atlas CT to patient simulation CT and obtain the liver’s Couinaud segments for the entire patient cohort. Secondly, a new attention convolutional neural network has been applied to utilize multimodality images (CT, MR and 3D dose distribution) for the prediction of high-risk segments. The results showed much improved efficiency for obtaining segments compared with conventional registration methods and the prediction performance showed promising accuracy for anticipating the recurrence location as well. Overall, this thesis contributed new methods and techniques to improve the utilization of radiomics for personalized radiotherapy. These contributions included new algorithm for detecting artifacts, a joint model of dose with image heterogeneity, combining hand-crafted features with machine learnt features for actuarial radiomics modeling, and a novel approach for predicting location of treatment failure.