IDH and 1p19q Diagnosis in Diffuse Glioma from Preoperative MRI Using Artificial Intelligence

Abstract

AbstractBackgroundIsocitrate dehydrogenase (IDH) mutation and 1p19q codeletion are important beneficial prognosticators in glioma. IDH and 1p19q diagnosis requires tissue sampling and there are likely benefits of presurgical diagnosis. Research supports the potential of MRI-based IDH and 1p19q diagnosis, however there is a paucity of external validation outside the widely used The Cancer Imaging Archive (TCIA) dataset. We present a combined IDH and 1p19q classification algorithm and assess performance on a local retrospective cohort (NZ) and the Erasmus Glioma Database (EGD).Methods2D convolutional neural networks are trained to provide IDH and 1p19q classification. Inputs are T1 post-contrast, T2, and FLAIR sequences. Training data consists of preoperative imaging from the TCIA dataset (n=184) and a locally obtained NZ dataset (n=349). Evaluation data consists of the most recent cases from the NZ dataset (n=205) and the EGD (n=420).ResultsIDH classification accuracy was 93.3% and 91.5% on the NZ and EDG, with AUC values of 95.4% and 95.8%, respectively. 1p19q accuracy was 94.5% and 87.5% with AUC values of 92.5% and 85.4% on the NZ and EGD datasets. Combined IDH and 1p19q accuracy was 90.4% and 84.3% on the NZ and EGD, with AUC values of 92.4% and 91.2%.ConclusionsHigh IDH and 1p19q classification performance was achieved on the NZ retrospective cohort. Performance generalised to the EGD demonstrating the potential for clinical translation. This method makes use of readily available imaging and has high potential impact in glioma diagnostics.Key Points-IDH and 1p19q are the main molecular markers in glioma.-Accurate predictions can be obtained from preoperative MRI without changes to imaging protocols.-Non-invasive diagnosis will likely enhance treatment planning and facilitate targeted preoperative therapies.Importance of the StudyThe 2021 WHO CNS tumour classification system formalises the increasing recognition of molecular factors like IDH and 1p19q in the prognostication and treatment of glioma. Emerging research shows the potential of artificial intelligence methods applied to preoperative MRI sequences to noninvasively predict molecular status. A limitation of the literature published to date is a lack of generalisation and external validation outside the widely used TCIA dataset. Here we present the performance of an MRI-based IDH and 1p19q classification tool evaluated on a large consecutive cohort from New Zealand and an independent publicly available dataset of MR images from the Netherlands. We demonstrate high predictive performance with robust generalisation, indicating the potential usefulness of this method in the workup of glioma. Reliable preoperative tumour characterisation may facilitate tailored treatment approaches and early decision making without the need for additional imaging.