Artificial intelligence (AI) models can effectively characterise brain tumour imaging data in real world situations, and spot tumour features on MRI scans previously detectable only with the use of special imaging techniques, according to new UCL research.
It means the technology could help ensure patients receive personalised and objectively evidenced care.
Analysing the properties of brain tumours in detail is essential for patient care. AI tools can do this well in research settings, where the imaging data they work with is high quality. But until now the ability of these AI models to deal with the ‘noisy’ and incomplete data characteristic of real world situations has not been known.
In the largest and most comprehensive study of its kind, the research team at UCL Queen Square Institute of Neurology applied state-of-the-art tumour imaging models to large-scale data from 1251 patients across multiple hospital sites.
The team looked at how precisely these AI models were able to evaluate different features of the brain found on MRI scans – in situations where the imaging data used was of high quality, low quality, and varying levels of quality in between.
The researchers found that models reliant on incomplete sets of MRI data were still able to characterise brain tumours well, and could even identify components of tumours that normally require special imaging techniques involving the use of injectable contrast agents. When injected into the body, these agents reveal distinctive tumour characteristics of clinical significance.
The work shows that the revolution in AI in research settings is readily translatable to real-world clinical settings in this area of radiology.
Lead author of the paper, Dr James Ruffle, said: “The path of translating research technologies to the real world is usually longer and more tortuous than it seems. But here we show it to be surprisingly direct, in the process revealing new avenues – the extraction of human unreadable signals – with great potential clinical value.”
The team’s work was funded by the Wellcome Trust, the NIHR UCLH Biomedical Research Centre, the Medical Research Council, and the Guarantors of Brain.
Read the paper – published in Brain Communications – in full.