Glioblastoma (GBM) remains the most aggressive cancer of the brain. Typical survival in patients with GBM is around 12-18 months, due to local or distant recurrence that occurs even after treatment with radiation and chemotherapy. Although there have been improvements in modern imaging, the radiation planning protocols are still based purely on a 2 cm geometric expansion of conventional post-contrast T1-weighted and T2-weighted FLAIR anatomic sequences. As a result, about 60% of tissue within the high dose treatment field can be normal brain tissue, which can damage to healthy brain function, while microscopic disease farther from the primary tumor bed is untreated. The main goal of this project was to develop a pipeline for the integration of probability maps derived from metabolic and diffusion-weighted MRI into the clinical workflow for radiation treatment planning. 24 patients with newly-diagnosed glioblastoma, who had undergone RT and chemotherapy consisting of an anti-angiogenic agent, were scanned at baseline prior to therapy and had serial follow-up imaging every 2 months until progression. Four patients were excluded because of either a poor initial model fit that resulted in inaccurate probability maps or poor image quality at the time of progression, leaving a total of 20 patients for evaluating our automatic contouring algorithm. First, we determined the optimal threshold for each patient’s probability map based on ROC analysis of the overlap of probability map with the progressed lesion. This value was then projected back on the histogram to automatically calculate each patient’s threshold based on the individual patient’s histogram and a maximum distance cutoff of 5 cm based on standard clinical procedures of high dose delivery during RT planning. Our results show that we were able to develop an automated contouring routine for integration of metabolic and physiology imaging into clinical workflow for radiation treatment planning. Incorporating a maximum distance from the original lesion allowed the automatic selection of a threshold from a consistent position on the histogram of the probability maps that optimized the overlap with the progressed lesion.

Gliomas are highly infiltrative, heterogenous brain tumors with poorly defined margin, and varying overall survival based on molecular subtype and grade. Despite recent developments in new diagnostic and treatment tools for gliomas, progression free survival and overall survival has only improved marginally for patients with glioma. Furthermore, treatment of glioma tends to be “one size fits all”, which can lead to either undertreating or overtreating the subclinical disease. Thus, the management of gliomas needs to be more patient-specific and more flexible over the course of the disease if the goal is to maximize both the longevity and quality of life of these patients.Recent advancement in MRI and radiation therapy research has opened the door for many opportunities to answer these questions. While the use of MRI in the clinic has been mostly limited to anatomical imaging, other MRI modalities have been gaining a lot of traction and have been proven to be able to provide clinical information not available in anatomical MRI. However, incorporating multimodal MRI in glioma management is a difficult task, because more advanced MRI acquisitions are not consistently acquired across institutions, and effectively understanding the consequences of changes observed on multimodal MRI over time is difficult even for trained radiologist. Artificial Intelligence has shown promise in making predictions from multi-parametric images, as multiple inputs can be given at the same time, and all processing and prediction tasks can be pre-trained and automatic applied. In this dissertation, we attempted to use multimodal MRI and artificial intelligence to improve both the diagnosis and treatment planning for patients with glioma. First, we developed an 1D deep learning based model that can help predict tumor histopathology noninvasively using the full spectrum of 1H MR Spectroscopic Imaging data. Then, we developed 3D segmentation-based deep learning model using multi-parametric MRI to redefine the clinical target volume for radiotherapy treatment planning and found that our model performed better than current practice, both in terms of better detecting subclinical disease and future progression, as well as sparing normal brain tissue. Finally, we highlighted the efficacy of using multi-parametric MRI in predicting a patient’s progression free and overall survival and improved the model performance by applying different types of masks.

Current methods of glioma pathology assessment using the tumor score metric rely on the extraction of a biopsy sample for evaluation by a pathologist. This method is limited by the fact that tumor score can vary within a glioma and that it only gives information regarding glioma pathology at one time point. An approach in which allows for the assessment of glioma pathology at various timepoints and in the entire brain is thus desirable. We explored such a method of glioma pathology prediction with machine learning, using both traditional and deep learning approaches. Using a dataset of patient information (MRI images and corresponding tumor scores, we performed several experiments with traditional machine learning models to explore the potential benefits of a deep learning based approach. We then developed, trained, and tuned a deep learning model that predicted tumor score from MRI data, and experimented with various forms of transfer learning to evaluate the impact of loading weights from different autoencoders. We determined the results of our traditional machine learning experiments showed a potential for a deep learning model’s ability to predict tumor score from MRI data. When evaluating our deep learning model we found that domain shift played a significant role in affecting our results in terms of testing accuracy, and we explored several methods to alleviate this issue. That said, our deep learning approach did outperform our traditional machine learning models, indicating the effectiveness of this approach.

Magnetic resonance imaging (MRI) is a high-resolution, non-invasive medical imaging modality that is widely used in human brain. In recent years, susceptibility weighted imaging (SWI) and quantitative susceptibility mapping (QSM) have been proposed to utilize MR phase signal to generate contrast from tissue magnetic susceptibility and even quantify the property. On the other hand, deep learning, especially deep convolutional neural networks (DCNNs), have achieved state-of-the-art performances in numerous computer vision tasks and gained significant attention in the field of medical imaging in the recent years. This dissertation combined the idea of deep learning with the two MR phase imaging methods.

To combined deep learning with SWI, we designed and trained a 3D deep residual network that can distinguish false positive detected candidates from cerebral microbleeds (CMBs) and built an automatic CMB detection pipeline with high performance. We further confirmed the generalizability of this deep learning-based pipeline using multiple dataset with different scan parameters and pathologies and provided lessons for application and generalization of generic deep learning based medical imaging methods.

To combine deep learning with QSM, we developed a 3D U-Net based network that learns to perform dipole inversion from gold standard QSM acquired from data with multiple orientation. The model was further improved with adversarial training strategy and achieved significantly lower reconstruction error than traditional QSM algorithms. In addition, we also performed various background removal and dipole inversion algorithms on both brain tumor patients and healthy volunteers to study and compare their performances. The results could provide guidance on future application of QSM in different scenarios.

Glioma is a heterogeneous and incurable neoplastic mass derived from the glial cells in the brain. The clinical course of a glial tumor can range from slow growing to highly aggressive and it is driven by the genetic and epigenetic alterations within the neoplastic cells’ DNA. In order to diagnose both that a patient has a glioma and what kind of glioma it may be, it is necessary to acquire magnetic resonance images (MRI) of the brain. MRI not only provides unparalleled soft tissue contrast in the brain compared with other tomographic imaging techniques (e.g. CT), it is also incredibly flexible: in addition to structural anatomy of the lesion, it can probe the physiology (e.g. diffusion, perfusion) and metabolism of the brain. Together, these data guide neuroradiologists and neurooncologists to provide the optimal treatment plan for a patient with glioma. Even with the most talented clinicians and the highest-quality MR acquisitions, there still exist issues along the trajectory of diagnosing a glioma. In this dissertation, I harnessed the incredibly rich biological information within the pixels of MRI with modern advances in data science and computer vision to address three of the most urgent problems along the diagnostic pipeline: 1) Can we identify a patient’s glioma subtype prior to surgical intervention?; 2) Can we create an automatic MR dashboard for clinicians to monitor patient disease over time?; and 3) When a patient appears to recur, is it a true glioma or the effect of radiation therapy?

The use of radiation therapy for treatment of brain tumor is controversial due to its long-term effects on neurocognitive function, especially in pediatric patients. Developing objective criteria for evaluating the severity of radiation-related injury is critical for children, due to their longer survival times and impact on early cognitive development. Cerebral microbleeds (CMB) which are deposits of hemosiderin that initially accumulate around vessels and can appear as early as 6 months post radiation therapy and continue to increase in number over time, however their vascular etiology is unknown. The aim of this project is to develop a method for simultaneous visualization of arteries, veins, and CMBs in order to automatically calculate vascular metrics from the fusion of MRA and SWI images obtained from a multi echo sequence at 7 Tesla. A strategy to assess the distribution of CMB’s relative to surrounding arteries and veins would help establish a connection between CMB formation and underlying vascular pathology. The tools developed in this study will be evaluated in pediatric patients with CMBs who have been treated with uniform supratentorial cranial radiation therapy for childhood brain tumors in order to ultimately determine whether metrics describing vascular structure could serve as quantitative markers for radiation injury, help in identifying regions of the brain that are most susceptible to radiation, and predict the formation of new CMBs and subsequent cognitive impairment.

Current encoder-decoder convolutional neural networks (CNN) used in automated glioma lesion segmentation and volumetric measurements perform well on newly diagnosed lesions that have not received any treatment. However, there are challenges in generalizability for patients after treatment, including at the time of suspected recurrence. This results in decreased translation to clinical use in the post-treatment setting where it is needed the most. A potential reason is that these deep learning models are primarily trained on a singular curated dataset and demonstrate decreased performance when they are tested in situations with unseen variations to disease states, scanning protocols or equipment, and operators. While using a highly curated dataset does have the benefit of standardizing comparison of models, it comes with some significant drawbacks to generalizability. The primary source of images used to train current models for glioma segmentation is the BraTS (Multimodal Brain Tumor Image Segmentation Benchmark) dataset. The image domain of the BraTS dataset is large, including high- and low-grade tumors, varying acquisition resolution, and scans from multi-center studies. Despite this, it may still lack sufficient feature representation in the target clinical imaging domain. Here we address generalizability to the disease state of post-treatment glioma. The current BraTS dataset consists entirely of images obtained from newly diagnosed patients who have not undergone surgical resection, received adjuvant treatment, or shown significant disease progression, all of which can greatly alter the characteristics of these lesions. To improve the clinical utility of deep learning models for glioma segmentation, they must accommodate variations in signal intensity that may arise as a result of resection, tissue damage (treatment induced or otherwise), or progression. We compared models trained on either BraTS data, UCSF acquired post-treatment glioma data, UCSF acquired newly diagnosed glioma data, and various combinations of these data, to determine the effect of including images with features unique to treated gliomas into training the networks on segmentation performance in the post-treatment domain. Although an absolute threshold training inclusion value for generalization of segmentation networks to post-treatment glioma patients has not been established, we found that with 200 total training volumes, models trained with greater than or equal to 30% of the training images from patients with prior treatment received the greatest performance gains when testing in this domain. Additionally, we found that after this threshold is met, additional images from newly diagnosed patients did not negatively impact segmentation performance on patients with treated gliomas. We also developed a pre-processing pipeline and implemented a loss penalty term that incorporates cavity distance relationships to the tumor into weighting a cross entropy loss term. The aim of this was to bias the network weights to morphological features of the image relevant to pathologies that are prevalent post-treatment. This may either be used as an initialization for training with an available larger dataset such as BraTS or used to finetune a transferred network that has not seen sufficient post-treatment glioma images during training in order to allow domain adaptation with fewer training data from this disease state. Preliminary results show qualitatively more desirable segmentations of tumor lesions with respect to cavities and small disconnected components in selected examples that are worthy of further analysis with alternate training configurations, more focused performance assessments, and larger cohorts. Here, we will evaluate these techniques as potential solutions to improve the generalizability of CNN tumor segmentation to post- treatment glioma, as well as provide a framework for further data augmentation based on augmenting the boundary of these lesions.