A study from UT Southwestern researchers shows that artificial intelligence can identify a specific genetic mutation in a glioma tumor simply by examining 3D images of the brain—with more than 97% accuracy.
Such technology could potentially eliminate the common practice of pretreatment surgeries in which glioma samples are taken and analyzed to choose an appropriate therapy.
“Knowing a particular mutation status in gliomas is important in determining prognosis and treatment strategies,” Joseph Maldjian, chief of neuroradiology at UT Southwestern’s O’Donnell Brain Institute, said in a statement. “The ability to determine this status using just conventional imaging and AI is a great leap forward.”
The study used a deep-learning network and standard MRI to detect the isocitrate dehydrogenase gene, which produces an enzyme that in mutated form may trigger tumor growth in the brain.
Doctors preparing to treat gliomas often have patients undergo surgery to obtain tumor tissue that is then analyzed to determine the IDH mutation status. The prognosis and treatment strategy will vary based on whether a patient has an IDH-mutated glioma.
However, because obtaining an adequate sample can sometimes be time-consuming and risky—particularly if tumors are difficult to access—researchers have been studying non-surgical strategies to identify IDH mutation status.
The study, published this spring in Neuro-Oncology, differentiates itself from previous research in three ways:
The method is highly accurate. Previous techniques have often failed to eclipse 90 percent accuracy.
Mutation status was determined by analyzing only a single series of MR images, as opposed to multiple image types.
A single algorithm was required to assess the IDH mutation status in the tumors. Other techniques have required either hand-drawn regions of interest or additional deep-learning models to first identify the boundaries of the tumor then detect potential mutations.
“The beauty of this new deep-learning model is its simplicity and high degree of accuracy,” Maldjian said. “We’ve removed additional pre-processing steps and created an ideal scenario for easily transitioning this into clinical care by using images that are routinely acquired.”
Maldjian’s team developed two deep-learning networks that analyzed imaging data from a publicly available database of more than 200 brain cancer patients from across the U.S.
One network used only one series from the MRI (T2-weighted images), while the other used multiple image types from the MRI. The two networks achieved nearly the same accuracy, suggesting that the process of detecting IDH mutations could be significantly streamlined by using only the T2-weighted images.
Maldjian’s team will next test his deep-learning model on larger datasets for additional validation before deciding whether to incorporate the technique into clinical care.
Meanwhile, researchers are hoping to develop medications to inhibit IDH through ongoing national clinical trials. If effective, these inhibitors could combine with AI-imaging techniques to overhaul how some brain cancers are assessed and treated.
