A recently published study by Karl Landsteiner University for Health Sciences demonstrated the use of artificial intelligence in brain tumour classification.
The study that was conducted by the Karl Landsteiner University for Health Sciences (KL Krems) found that multiclass machine learning methods could be used to analyse and classify brain tumours using physiological data from magnetic resonance imaging.
The results were then compared with classifications made by human experts. Artificial intelligence was found to be superior in the areas of accuracy, precision and misclassification, among others, while professionals performed better in sensitivity and specificity.
The team led by Professor Andreas Stadlbauer, a scientist at the Central Institute for Medical Radiology Diagnostics at St. Pölten University Hospital, used both advanced and physiological MRI data for the study. Both methods provide enhanced insight into the structure and metabolism of a brain tumour and have allowed better classification for some time. But the price to pay for such a differentiated picture is enormous amounts of data that need to be expertly assessed. ‘We have now analysed whether and how an artificial intelligence using ML can be enabled to support trained professionals in this Herculean task states Stadlbauer. ‘And the results are very promising. When it comes to accuracy, precision and avoiding misclassification, an AI can classify brain tumours well using MRI data.
Brain tumours can be easily detected by magnetic resonance imaging (MRI), but their exact classification is difficult in this way. Yet that's precisely what's crucial for choosing the best possible treatment options. Now, an international team led by KL Krems has used data from modern MRI methods as the basis for machine learning (ML) protocols and assessed the use of artificial intelligence to classify brain tumours. They found that in certain areas, classification using artificial intelligence can be superior to that performed by trained professionals.
To achieve their impressive result, the team trained nine well-known Multiclass ML algorithms with MRI data from 167 previous patients who had one of the five most common brain tumours and had accurate classification using histology. A total of 135 so-called classifiers were generated in a complex protocol. These are mathematical functions that assign the material to be examined to specific categories.
‘In contrast to previous studies, we also took into account data from physiological MRIs,’ explains Professor Stadlbauer. ‘This included details on the vascular architecture of the tumours and their formation of new vessels, as well as the supply of oxygen to the tumour tissue.’
The team named the combination of data from different MRI methods with multiclass ML ‘radiophysiomics.’ It's a term that's likely to catch on quickly, as the potential of this approach became apparent in the second part of the project, the testing phase. In this, the now-trained multiclass ML algorithms were fed with corresponding MRI data from 20 current brain tumour patients and the results of the classifications thus obtained were compared with those of two certified radiologists. Thereby, the two best ML algorithms (referred to as ‘adaptive boosting’ and ‘random forest’), outperformed the human assessment results in the areas of accuracy and precision. Also, these ML algorithms resulted in less misclassification than by the professionals (5 versus 6). On the other hand, when it came to the sensitivity and specificity of the assessment, the human assessments proved to be more accurate than the AI tested.
‘This also makes it clear,’ says Professor Stadlbauer, ‘that the ML approach should not be a substitute for classification by qualified personnel, but rather a supplement to it. In addition, the time and effort required for this approach is currently still very high. But it offers a possibility whose potential should be further pursued for everyday clinical use.’ Overall, this study again demonstrates the focus of research at KL Krems on fundamental findings with real clinical added value.
