Non-Invasive Molecular Status Differentiation of Pediatric Low-Grade Gliomas From Magnetic Resonance Images Using Machine Learning

Pediatric low-grade glioma (pLGG), the most prevalent central nervous system tumor in childhood, is often driven by one of two genetic alterations: BRAF Fusion or BRAF Mutation. Identification of genetic status is critical for optimal prognostication and treatment of pLGG. Genetic status is typically determined through biopsy, which has associated risks, and in some cases is not possible due to tumor location. A prior study showed that machine learning (ML) approaches could potentially detect pLGG genetic status non-invasively from MR images, reducing the need for surgery. Through three key studies, this thesis aimed to make progress towards a clinically deployable imaging-based pLGG genetic status prediction ML model. When published, the study featured in Chapter 2 provided the most robust evidence available on the ability of ML to distinguish between patients with pLGG driven by BRAF Fusion and Mutation using radiomic features, such as shape or texture, extracted from manually segmented tumor regions of MR images. Strong performance was achieved using a larger cohort and more robust statistical methods than earlier works, verifying the potential of ML-based BRAF status prediction. Additionally, to help encourage clinical uptake, an uncertainty quantification framework and nomogram were developed and validated. The study Chapter 3 was based on expanded beyond hand-crafted radiomics features to investigate an additional method of classifying medical images, deep learning (DL). This study focused on the more realistic task of grouping pLGGs into three categories: BRAF fusion, BRAF mutation, and non-BRAF altered, unlike most earlier studies which ignored cases in the latter class. A comprehensive ML framework combining DL and radiomics proved optimal, outperforming either approach on its own. Chapter 4 introduced a segmentation-free approach, eliminating dependency on manual or automated tumor segmentations. Instead, whole-brain MR images were directly analyzed. A novel in-domain pretraining regimen was used to build tumor segmentation information into classification models. The segmentation-free framework achieved classification accuracy similar to a manual segmentation-based pipeline, without requiring tumor segmentations as an input.Collectively, the findings of this thesis demonstrate the potential of ML to enable non-invasive pLGG genetic profiling, which could help improve outcomes for patients with pLGG.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.