Non-invasive MRI-based Pipelines for Pediatric Low-Grade Gliomas Molecular Biomarker Identification and Beyond Using Machine Learning

Pediatric low-grade glioma (pLGG) is the most common type of brain cancer among children, and molecular marker identification of pLGG is crucial for successful treatment planning. The current standard of care is biopsy, which is invasive. Thus, non-invasive imaging-based approaches, where Machine Learning (ML) has high potential, are impactful. We curated a large dataset of pretherapeutic Magnetic Resonance Imaging (MRI) scans for developing non-invasive pLGG molecular subtype identification. In this work, first, we conducted dataset size sensitivity analysis and determined that 132 patients would suffice for training a reliable pLGG molecular subtype identification pipeline using MRI-based radiomics. Additionally, we observed performance variance when training and validation splits were repeated. Inspired by the first study, we proposed Open-radiomics as a collection of standardized radiomics datasets, a radiomics research protocol, and a base pipeline for radiomics-based brain tumour pathology classification using open-source adult patient datasets. In the third study, we introduced the IMICS ROC analyzer, a method and software for the identification of outliers in retrospective studies. Relying on the first three studies, in the fourth study, we redesigned the pipelines to (1) provide per-patient ML-based inference for our bi-institutional pLGG radiomics dataset and (2) identify outliers using IMICS ROC analyzer. The open-radiomics-based design of the fourth study enabled us to measure the confidence of our subtype identifiers using a Monte Carlo method. We included three additional studies using more advanced ML techniques to improve the pipelines. In the fifth study, we demonstrated that Large Language Models (LLMs) could augment the clinical data and improve the pLGG subtype identification pipelines. In the sixth study, we used convolutional neural networks (CNN) and proposed 3D probability distribution functions of tumour location to allow the classifiers to explore beyond the manual segmentation masks provided by neuroradiologists. In the last study, we improved the performance of the CNNs using a novel multiclass area under the receiver operating curve (AUROC) loss function. We leveraged radiomics, ML, LLMs, and CNNs and developed non-invasive pLGG molecular subtype identification pipelines tailored for precision medicine, enabling improved clinical adoption and patient outcomes.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.