Thursday, December 5 | 2:20 p.m.-2:30 p.m. | R6-SSNR16-6 | Room S406B
AI shows promise for characterizing brain tumor subtypes found on preoperative MR imaging, according to research to be shared with attendees on Thursday afternoon.
The use of AI in this manner could offer a noninvasive way to facilitate preoperative planning and postoperative treatment, lead author Hugh McHugh, MBChB, of Vancouver General Hospital in Canada and colleagues noted.
Molecular marker prediction from preoperative MRI is an "appealing application of AI in radiology," the group wrote in an abstract about their research. "[Status of the markers] isocitrate dehydrogenase (IDH) and 1p19q … determine whether a glioma is a glioblastoma, astrocytoma or oligodendroglioma, which has important prognostic implications."
Tissue analysis requires surgery, so finding a noninvasive alternative would be a boon for patients, the investigators noted. But an alternative to surgery must be able to differentiate brain cancer from other disease, in particular central nervous system lymphoma and metastasis. McHugh and colleagues developed a 2D convolutional neural network to predict tumor histology as well as IDH and 1p19q status and used it in a study that included 935 patients with brain tumors. Of the study participants, 557 made up a training cohort and 378 a testing cohort. Patients underwent MRI with T1 postcontrast, T2, and FLAIR sequences; tumor classes were glioblastoma, astrocytoma, oligodendroglioma, central nervous system lymphoma, and metastasis.
The team found the following:
- The accuracy of the algorithm for predicting glioma versus other tumors was 84.2%. Sensitivity was 84.4%, specificity was 83.8%, and the area under the receiver operating curve (AUC) was 90.9%.
- The algorithm's accuracy for predicting IDH status was 95%; its sensitivity was 53.3%, its specificity was 98.6%, and the AUC was 95.2%.
- The algorithm's accuracy for predicting 1p19q status was 96.8%, while its sensitivity was 50% and its specificity, 98.6%. The AUC was 93.9%.
The algorithm's performance for identifying glioma versus non-glioma was high -- an important quality for the purpose of predicting glioma molecular subtype.
"Further work will incorporate additional sequences such as diffusion-weighted imaging [DWI] to help differentiate non-gliomas, followed by a multicenter prospective study," the authors concluded.
Get more information about this study by attending this session.
