Identifying glioma biomarkers takes a biopsy (for now), but earlier access to that info is on the way.
MRIs can identify key biomarkers for glioma and predict tumour histology, providing vital information previously unavailable before brain surgery takes place.
Artificial intelligence can now classify key biomarkers for glioma using an algorithm developed by radiology registrar Dr Hugh McHugh, who presented the data at the 2022 RANZCR New Zealand Annual Scientific Meeting in Queenstown earlier this month.
IDH mutation and 1p19q codeletion are the key determining factors for glioblastoma (IDH wild-type), astrocytoma (IDH mutant-1p19q intact) and oligodendroglioma (IDH mutated-1p19q co-deleted). So far, it’s only been possible to identify them in patients through a brain biopsy.
But knowledge of their presence could help determine the best surgical approach for that patient. Diffuse glioma patients with IDH1 mutation have better survival rates than IDH1 wild-type tumours after surgical resection. And Dr McHugh told the conference that those who have both markers fare much better.
“Patients [in the study cohort] with IDH wild type tumours lived for about a year. With IDH mutated 1p19q intact, they lived for 12 years, so quite a significant jump. And for IDH mutant-1p19q co-deleted, the median wasn’t defined with more than half of these patients being alive at the time of data collection. Estimated mean survival for these patients was 16 years,” he said.
Dr McHugh developed a model enabling AI to interpret visual data from MRI and correctly classify these biomarkers. He programmed the 2D convolutional neural networks needed to automate this process by using pre-operative images for adult patients with grade II to IV glioma from The Cancer Imaging Archive (184 cases) and 349 cases from a local New Zealand dataset.
“The output is a segmentation which provides the classification,” Dr McHugh explained.
“Basically, it’s an overlay of where the tumour is, blue being GBM, red being astrocytoma, and green being oligodendroglioma. You also get tumour volumes, which is useful, but the main ticket is the mutation status.”
The algorithm was tested on 250 cases from the New Zealand data set (from 2018-2021, rather than those used for training purposes, which were from 2011-2017) and 420 cases from the Dutch Erasmus Glioma Database (EGD).
Sensitivity and specificity were high for both datasets, as shown by the AUC (area under curve) value, which measures model accuracy across a range of sensitivity/specificity thresholds.
For IDH, classification was 93.3% correct in the NZ cohort (AUC value 95.4%) and 91.5% for EDG (AUC of 95.8%). For 1p19q, the AI was correct 94.5% of the time in the NZ data set (AUC 92.5%), and 87.5% in the EGD (AUC 85.4%). Combined accuracy was 90.4% (AUC 92.4%) in NZ, and 84.3% (AUC 91.2%) in the EGD dataset.
The software provided an automated quality control report for each case, said Dr McHugh.
“So, by looking at an individual case you can get an indication of how certain the model is that you’ve got a specific tumour.
“The specificity is much better than the sensitivity and this is useful because it’s probably more important to be sure that you have one of the more favourable tumours rather than the other way around, particularly if you’re going to be changing management on that information,” he said.
Dr McHugh told Oncology Republic the first step towards bringing the technology into clinicians’ offices was to integrate the method into a prospective study over the next 12 months as he completed his radiology specialist training.
“It would be great to have some software up and running here in New Zealand in the next couple of years,” he said.
“The plan is to include brain metastases and CNS lymphomas into the analysis as these tumours have similar radiological appearances to glioma. So being able to differentiate between these tumours would be very useful.”
Dr Hugh McHugh is the winner of the OBEX Medical Registrar Research Award for the best Radiology Registrar Presentation at this year’s NZ ASM.
