Fingerprinting Heterogeneity of Glioma using PET-MR Information
Session type: Poster / e-Poster / Silent Theatre session
Theme: Diagnosis and therapy
Gliomas are considered the most frequent brain tumors with aggressive behavior leading and a high mortality rate. They are usually associated with a high degree of intra-tumor heterogeneity. For the personalized glioma diagnosis and therapy, it is critical to characterize different parts of glioma with individual malignancy for biopsy and therapy guidance. The era of PET/MRI provides the opportunity to capture different anatomical and molecular perspective of glioma. However, it is still very challenging to properly identifying individual parts of glioma from multi-parametric images acquired in PET/MRI.
This thesis is proposing a novel machine learning algorithm based on Generative Method to characterize intra-tumor heterogeneity of glioma. The algorithm was applied on dynamic [18F] FET-PET, [18F] Fmiso PET, rOEF, MRI T1, T2, T1W, T2W, FLAIR, DCE MRI and so on. This probabilistic model allows for different tumor boundaries in each channel, reflecting difference in tumor appearance across modalities.
Magnetic resonance imaging (MRI), as the gold standard diagnostic tool for brain tumors, offers high spatial resolution and is widely available. In high-grade gliomas (HGG) the area of contrast enhancement on MRI T1-weighted sequences is generally assumed to reflect the main tumor burden. In neuropathologic studies, however, invasive glioma cells can be found far beyond contrast enhancing areas. Recently, molecular imaging studies using the [18F]-FET tracer revealed that in HGG patients the “metabolic tumor volumes” are subject and instrument dependent which brought us about this conclusion into four classes.
We have developed a model-based segmentation method for segmenting head PET/MR image datasets with tumors. This is achieved by extending the spatial prior of a statistical normal human brain atlas with individual information derived from the patient’s dataset. Thus, we combine the statistical geometric prior with image-specific information for both geometry of newly appearing objects, and probability density functions for healthy tissue and pathology.