Dissertation Date: June 25, 2025
Approximately 25,000 newly diagnosed malignant brain and central nervous system (CNS) tumors are reported in the US annually. Approximately 80% of these cases are gliomas and the most common and aggressive type is histologic glioblastoma multiforme (GBM) with a median overall survival of 12-14 months despite intensive treatment regimens including maximal safe resection followed by combination radiation therapy and chemotherapy. Unfortunately, these strategies are generally not curative, a result largely attributed to the infiltration of GBM cells into surrounding tissue, inevitably leading to recurrence. This is reflected in research indicating survival benefits may be gained from supramaximal resection of both T1-weighted contrast-enhancing lesion (CEL) and a portion of surrounding non-contrast enhancing, FLAIR (Fluid Attenuated Inversion Recovery) hyperintense lesion (NEL). However, there remains a need for an imaging biomarker capable of distinguishing tumor cell infiltration from vasogenic edema within NEL, which would enable optimized treatment planning with precise surgical and focused radiation targets and may prove useful during treatment surveillance, providing a more complete assessment of tumor response or recurrence.
Although existing studies have reported promising results using machine learning in combination with multiparametric MRI (mpMRI) to address this need, their interpretability remains limited due to suboptimal data acquisition and methodological constraints, including small cohort sizes, unreliable tissue labeling, and insufficient validation procedures.
The primary objective of this study was to develop and train a convolutional neural network (CNN), termed the infiltrative tumor burden (iTB) model, to differentiate infiltrative tumor tissue from peritumoral edema within the non-contrast enhancing lesion (NEL) region of gliomas, utilizing multiparametric magnetic resonance imaging (mpMRI) data. The model was trained on a large, histologically validated dataset comprising stereotactically acquired biopsy samples that were spatially co-registered with mpMRI, providing a high-fidelity ground truth for tissue classification and enabling robust evaluation on an independent test set. The mpMRI inputs chosen for this study were T1, T1+C, T2, FLAIR, ADC (apparent diffusion coefficient), rCBV (relative cerebral blood volume) and rCBF (relative cerebral blood flow). The CNN architecture was based on a 3D ResNet framework, incorporating ConvNeXt-inspired enhancements alongside additional architectural modifications tailored to the specific characteristics of the dataset and the study’s experimental requirements. The second objective of this study was to optimize the data augmentation process using a novel forward selection strategy, whereby the parameter controlling the applied magnitude of each augmentation technique is incrementally increased until a decline in model performance is observed. The
final objective of this study was to investigate key, clinically relevant properties of the iTB model by assessing its repeatability over short time intervals and conducting an input removal sensitivity analysis to elucidate how the model leverages mpMRI inputs to generate predictions.
The results demonstrate that the iTB model achieves performance comparable to or exceeding that of prior approaches, with further improvements observed following optimization of data augmentation strategies. Moreover, the resulting iTB maps were shown to be highly repeatable over short time frames. Finally, the sensitivity analysis revealed that each of the selected mpMRI input modalities contributed meaningfully to the model’s predictive performance, underscoring their relevance in the context of infiltrative tumor burden estimation.
