Abstract
Qualitative imaging, primarily focusing on brain tumors’ genetic alterations, has gained traction since the introduction of molecular-based diagnosis of gliomas. This trend started with fine-tuning MRS for detecting intracellular 2HG in IDH-mutant astrocytomas and further expanded into a novel research field named “radiomics”. Along with the explosive development of machine learning algorithms, radiomics became one of the most competitive research fields in neuro-oncology. However, one should be cautious in interpreting research achievements produced by radiomics as there is no “standard” set in this novel research field. For example, the method used for image feature extraction is different from research to research, and some utilize machine learning for image feature extraction while others do not. Furthermore, the types of images used for input vary among various research. Some restrict data input only for conventional anatomical MRI, while others could include diffusion-weighted or even perfusion-weighted images. Taken together, however, previous reports seem to support the conclusion that IDH mutation status can be predicted with 80 to 90% accuracy for lower-grade gliomas. In contrast, the prediction of MGMT promoter methylation status for glioblastoma is exceptionally challenging. Although we can see sound improvements in radiomics, there is still no clue when the daily clinical practice can incorporate this novel technology. Difficulty in generalizing the acquired prediction model to the external cohort is the major challenge in radiomics. This problem may derive from the fact that radiomics requires normalization of qualitative MR images to semi-quantitative images. Introducing “true” quantitative MR images to radiomics may be a key solution to this inherent problem.
Applying machine learning and image feature extraction techniques to the problem of cerebral aneurysm rupture