Machine learning improves early detection of glioma mutations

Machine learning (ML) methods can quickly and accurately diagnose mutations in gliomas, primary brain tumors.

This is supported by a recent study conducted by the Karl Landsteiner University of Medical Sciences (KL Krems). In this study, physiometabolic magnetic resonance imaging (MRI) data were analyzed using ML methods to identify mutations in a metabolic gene. Mutations in this gene have a significant impact on the course of the disease, and early diagnosis is important for treatment. The study also shows that there are currently inconsistent standards for obtaining physiometabolic MRI images, which hinders routine clinical use of the method.

Gliomas are the most common primary brain tumors. Although their prognosis is still poor, personalized therapies can significantly improve treatment success. However, the use of such advanced therapies relies on individual tumor data, which is difficult to obtain for gliomas due to their location in the brain. Imaging methods such as magnetic resonance imaging (MRI) can provide such data, but their analysis is complex, labor-intensive, and time-consuming. The Central Institute for Diagnostic Medical Radiology at the University Hospital St. Pölten, the teaching and research base of KL Krems, has been developing machine and deep learning methods for many years to automate such analyses and integrate them into routine clinical procedures. Now another breakthrough has been achieved.

"Patients whose glioma cells carry a mutated form of the isocitrate dehydrogenase (IDH) gene actually have a better clinical outlook than those with the wild type," explains Professor Andreas Stadlbauer, a medical physicist at the Zentralinstitut. "This means that the earlier we know the mutation status, the better we can individualise the treatment." Differences in the energy metabolism of mutated and wild-type tumours help in this. Thanks to previous work by Professor Stadlbauer's team, these can be easily measured using physiometabolic MRI, even without tissue samples. However, analysing and evaluating the data is a very complex and time-consuming process that is difficult to integrate into clinical practice, especially since results are needed quickly due to the poor prognosis of patients.

In the current study, the team used ML methods to analyze and interpret this data in order to obtain results faster and be able to initiate appropriate treatment steps. But how accurate are the results? To assess this, the study first used data from 182 patients from the University Hospital St. Pölten, whose MRI data were collected according to standardized protocols.

"When we saw the results of our ML algorithms," explains Professor Stadlbauer, "we were very pleased. We achieved an accuracy of 91.7% and a precision of 87.5% in distinguishing between tumours with the wild type of the gene and those with the mutated form. We then compared these values with ML analyses of classical clinical MRI data and were able to show that using physiometabolic MRI data as a basis gave significantly better results."

However, this superiority only held when analyzing data collected in St. Pölten using a standardized protocol. This was not the case when the ML method was applied to external data, i.e. MRI data from other hospital databases. In this situation, the ML method trained on classic clinical MRI data was more successful.

The reason why the ML analysis of physiometabolic MRI data showed worse results is that the technology is still young and in the experimental stage of development. Data collection methods still vary from hospital to hospital, which leads to biases in ML analysis.

For the scientist, the problem is "only" one of standardization, which will inevitably arise with the increasing use of physiometabolic MRI in different hospitals. The method itself - rapid assessment of physiometabolic MRI data using ML methods - has shown excellent results. Therefore, it is an excellent approach for determining the IDH mutation status of glioma patients before surgery and for individualizing treatment options.

The results of the study were published in the journal Karl Landsteiner University of Health Sciences (KL Krems).