Date of Award
Doctor of Philosophy (PhD)
According to 2016 cancer statistics, brain tumors are the leading cause of cancer-related morbidity and mortality around the world. More than 100 types of brain tumors, distinguishable by unique histopathological features, have been identified that differ significantly in prognosis and treatment strategies. Currently, histopathology is the diagnostic standard for characterizing brain tumors, which carries risks and potential complications. So, MRI is frequently used as an alternative to or in conjunction with histopathology due to its non-invasive nature and high soft-tissue contrast. With the emergence of artificial intelligence (AI) based approaches, different machine learning (ML) and deep learning (DL) models have been proposed for classification of tumors from MRI. Nevertheless, these methods are limited by reliance on manually engineered features, small patient cohorts used for training, and especially the requirement of manually segmented tumor volumes. Moreover, gliomas, the most common and aggressive malignant adult brain tumor, requires information of molecular parameters in addition to histopathology for a comprehensive classification. Currently, the clinical gold-standard of subtyping glioma involves invasive brain biopsy procedures that can be risky, may fail to capture intra-tumoral spatial heterogeneity due to localized samples, or can be inaccessible in low-resource settings. Multiple ML and DL models that have been proposed to classify molecular parameters by leveraging the variation in phenotypical characteristics of tumor manifested in MRI scans, are limited by the requirement of costly manual annotations, lack of rigorous validation, or applicability on only specific grades of glioma which hinders their translation to a pre-operative clinical setting. Additionally, the success of most existing AI-based methods proposed for tumor diagnosis are contingent on careful manual selection and pre-processing of MRI scans with appropriate tissue contrast properties, which is extremely time-intensive due to the high degree of non-uniformity in clinical neuro-oncology studies.Hence, the over-arching goal of my research is to build AI-driven solutions for curation, pre-processing, classification, and segmentation of intracranial tumors without the requirement of expert supervision. To this end, first, I have developed a 3d convolutional neural network (CNN) for classifying MRI scans into a healthy class and six most commonly occurring intracranial tumor classes viz., high grade glioma, low grade glioma, brain metastases, meningioma, pituitary adenoma, and acoustic neuroma, using only a single 3d post-contrast T1-weighted MRI volume per subject and without the requirement of any additional manual interaction. Second, I have built a hybrid CNN to simultaneously detect and segment glioma from MRI scans as well as classify two important molecular markers viz. the mutation of isocitrate dehydrogenase (IDH) enzyme and co-deletion status of chromosome arms 1p and 19q (1p/19q). The classification is performed by leveraging both MR imaging features and prior knowledge features acquired from clinical records and anatomical location of tumors. Third, I have designed and developed an end-to-end AI-driven framework for the aggregation and processing of multisequence neuro-oncology MRI data to extract quantitative tumor measurements. This framework i) classifies MRI sequence types using an ensemble classifier, ii) preprocesses the data in a reproducible manner, iii) delineates tumor tissue subtypes using CNNs, and iv) extracts diverse radiomic features. In this emerging era of precision diagnostics, these developments demonstrate a high potential for integration as assistive tools into clinical workflows to support clinical management and drive personalized treatment planning for intracranial tumors.
Daniel S. Marcus Joseph O'Sullivan
Abhinav K. Jha, Joshua S. Shimony, Aristeidis Sotiras,
Available for download on Wednesday, May 15, 2024