A tool developed in Barcelona, based on artificial intelligence, will improve the diagnosis of brain tumors. Researchers are thinking about the goal of being able to diagnose without having to resort to the currently almost inevitable biopsy, which in the case of the brain is a complicated surgical intervention.
The diagnosis of brain tumors is based on the evaluation of magnetic resonance images before and after administering contrast (substances that improve visualization or identify possible lesions in the examination). Although in most cases the radiological test is not enough to avoid subjecting the patient to a diagnostic neurosurgical procedure.
Discern, the AI designed by the radiomics group of the Vall d'Hebron Institute of Oncology (VHIO) in collaboration with researchers from the Bellvitge hospital, has learned to differentiate with a probability of success of 78% between the three most common brain types .
Glioblastoma multiforme, brain metastases from solid tumors and primary nervous system lymphoma represent 70% of malignant brain tumors and require different therapeutic approaches, making it essential to unequivocally identify them.
Researchers have taught AI what are the characteristics of tumors that appear on MRI scans from 50,000 voxels (the minimum processable unit in 3D MRI images, equivalent to a pixel on a 2D object) of 40 patients. already diagnosed. The tool has been validated in more than 500 additional cases and the machine is capable of correctly finding the repetitive patterns of each of the tumors in 78% of cases, a higher proportion than that obtained with current methods.
"One new thing about the project is that, instead of training the machine with the global information of the entire image, we have done it with the information of each of the voxels that make up each digital image of the tumor, so that we go from from a few hundred patients to thousands of sources for training,” explains Raquel Pérez-López, head of the VHIO radiomics group.
"We have asked him for a final diagnosis, but he also makes a diagnosis for each of the voxels in the image, so that when he gives the same diagnosis for 80% of the voxels the doctor can be more sure of the diagnosis," he adds. .
What happens in the cases (22%) of not getting it right? “In these cases, Discern gives the voxels diagnoses at random, so that if that happens the doctor can see that the machine is not clear and discard the diagnosis it gives,” clarifies the radiologist.
In short, the tool deeply analyzes the image data that the human eye is not able to perceive and finds the complex patterns that characterize one tumor compared to others. According to Pérez-López, the 78% success rate can be improved: “Experience in other projects indicates that when you train the machine with more cases, it improves.” At the same time, researchers are considering the idea of extending the selection to the rest of the brain tumors.
The results of the study have been published in the scientific journal 'Cell Reports Medicine' and the authors have developed open access software so that the tool can be used in any hospital to continue perfecting the diagnostic system.
Although it has been shown that the results are reproduced in other centers, with cases that the machine has not seen (in Andalusia and California) and exceed the diagnostic capacity of current tools, its application will not be immediate. “New biomarkers must go through a series of processes. You have to do a prospective clinical trial and these trials are costly in time and money,” explains the radiologist.
In his opinion, five years is a realistic period for Discern, after clinical trials that will last no less than two years, to be available. “The goal is for the tool to reach patients as soon as possible. It is an assistance to the diagnosis of the radiologist and the oncologist.” Optimally, it should contribute to avoiding, to a significant extent, the need for biopsies.
The Bellvitge and Clínic hospitals of Barcelona, HT Médica Andalucía and several departments of the University of California (USA) collaborate in the VHIO research. The CRIS Contra el Cancer, FERO and La Caixa foundations provide financial support.
