ASCO 2018: Advanced MRI—aided by AI—classifies brain tumors based on mutation status, decreasing diagnostic uncertainty

Researchers from the University College London and Image Analysis Group in the United Kingdom recently discovered that contrast perfusion-weighted MRI (DSC-PWI) enhanced by artificial intelligence (AI) and texture analysis can accurately and non-invasively differentiate between brain tumors according to mutation status. 

The researchers' findings were presented Monday, June 4, at the 2018 American Society of Clinical Oncology (ASCO) annual meeting in Chicago.  

A mutation in the encoding gene isocitrate-dehydrogenase (IDH) is defined as a molecular biomarker in diagnosing and classifying gliomas. The technique characterizes gliomas based on their isocitrate-dehydrogenase (IDH) mutation status obtained through analyzing tumor features from relative cerebral blood volume (rCBV) data maps.

Researchers—led by Sotirios Bisdas, MD, MSc, a consultant neuroradiologist at the University College London Hospitals NHS Foundation Trust and an associate professor of neuroradiology at University College London—created a non-invasive, advanced MRI decision-support tool to characterize and monitor tumors through machine learning visual analysis techniques. 

"The tool is about assurance and value—it helps decrease the uncertainty of diagnosis and increases or improves the therapeutic options," Bisdas told Health Imaging. "It allows, in the most basic way, to characterize the tumor and give this information to oncologists and we can use the same method for treatment monitoring. It's quite versatile, too. You can use it for any tumor."  

For their study, Bisdas and his team recruited 208 participants (57 percent male, with an average age of 47 years) diagnosed with gliomas from various health centers. Data from DSC-PWI were processed using a fully adaptive Bayesian method to create leakage-corrected rCBV maps where tumors were manually segmented and registered, the authors wrote.  

Next, rCBV maps were used to generate texture features over the tumor images. Machine learning was then used to analyze the extracted features and IDH statuses of each tumor. 

The researchers found that overall sensitivity and specificity rates for the rCBV for IDH stratification were 68 percent and 81 percent, also noting that "all except one of the 10 classical histogram statistics and 12 texture features appeared very different across mutation status when using non-parametric Wilcoxon test." 

In terms of classification across tumor grading, the same features led to a distance error less than or equal to one in 88.6 percent of the cases and an exact prediction in more than half of cases.  

"Stratifying of gliomas is important in terms of defining treatment strategies, accurate guidance for biopsy and also counseling in clinics," the authors wrote. "Heterogenous radiological findings presented by various immunohistopathological glioma subtypes may cause a challenge in conventional imaging, which can be augmented by advanced techniques, mainly perfusion studies, which are the most accurate tools in the estimation of neo-angiogenesis, the latter being an exact finding of high grade gliomas."  

Bisdas told Health Imaging he believes his team is aiming in the next year to advance their research by using deep learning algorithms and a larger participant population.