Deep Learning Approaches for Assisting MR-guided Radiation Therapy
Magnetic resonance-guided radiation therapy (MRgRT) has drawn enormous clinical and research interests. The superior soft-tissue contrast of magnetic resonance imaging (MRI) compared with computed tomography (CT) allows more accurate tumor and organ-at-risk (OAR) segmentation for brain, prostate, and abdominal cancer. Additionally, real-time target tracking ability and high-quality daily MR images offered by the online MRgRT system could further minimize treatment delivery uncertainties. However, the current MRgRT workflow has several limitations including the need to acquire an additional CT for treatment planning, slow tumor and OAR recontouring in the adaptive workflow, and underdeveloped tools for predicting treatment response and survival outcome. In this dissertation, we developed and investigated several deep learning (DL) methods to address these three limitations.
First, 2D and 3D convolutional neural networks (CNNs) were proposed to generate pelvic synthetic CT (sCT) images from 1.5T MR images. Second, conditional generative adversarial network (cGAN) and cycle-consistent generative adversarial network (cycleGAN) were investigated for abdominal sCT generation based on 0.35T MR images. Third, a novel multi-path 3D DenseNet was proposed for automatic glioblastoma multiforme (GBM) segmentation based on multi-modal MR images and compared with the corresponding single-path DenseNet. For predicting neoadjuvant chemoradiation treatment (nCRT) response in patients with locally advanced rectal cancer (LARC), two logistic regression models were built using handcrafted radiomic features and DL-based radiomic features, respectively. These radiomic features were extracted from pre-treatment diffusion-weighted MR images based on manually delineated gross tumor volume. Additionally, an automatic radiomic workflow was proposed for GBM survival prediction based on multi-modal MR images. This workflow consisted of an automatic tumor segmentation CNN and a Cox regression model.
The proposed 3D CNN generated more accurate pelvic sCT images compared with the 2D CNN. Abdominal sCT images generated by both GANs achieved accurate dose calculation for liver radiotherapy plans. The multi-path DenseNet achieved more accurate GBM segmentation compared with the single-path DenseNet. The logistic regression model constructed using DL-based features achieved significantly better classification performance in predicting nCRT response compared with the model constructed using handcrafted features. The proposed automatic workflow demonstrated the potential of improving patient stratification and survival prediction in GBM patients.
The proposed DL methods could potentially address three limitations of the MRgRT workflow but were investigated across different cancer types due to limited data availability. Future work could be adapting these methods for one cancer type and conducting further investigation to translate them into clinics.