Nimg-25. clinical feasibility of ai-based t2 flair segmentation for detection of progression in non-enhancing lower grade glioma

Abstract Radiological assessment of progression in patients with non-enhancing lower grade gliomas (LrGG: grade 2 and 3) is primarily based on increase in lesion size on anatomical T2-weighted fluid-attenuated inversion recovery (T2 FLAIR) magnetic resonance imaging (MRI) sequences. We evaluated whether an AI-based method could be used routinely as part of patient care to assist in detecting non-enhancing glioma progression. The software offers the ability for real-time quality control and editing, choice of baseline exam and threshold for progression, and shows 3D image segmentations side-by-side with lesion volume graphs. The model was trained to create T2 FLAIR abnormality segmentations. The 515 training images included pre-surgical scans of newly diagnosed and recurrent patients, and post-surgical surveillance scans of patients. Clara Deploy was used for inference and to handle communication between PACS and an XNAT server, where results were stored and displayed. Following RANO criteria, lesions were marked as ‘progressed’ if volume changes were larger than 40% between baseline and the MRI of interest. The 124 AI-based segmentation model test volumes were compared to manual segmentation volumes, and achieved a mean DICE score of 0.87 ± 0.20. The sensitivity and specificity of the RANO progression assessment based on the model was compared with RANO progression assessment based on manual segmentation in 16 patients (6 serial MRIs each), with 85% sensitivity, 82% specificity. This work presents the first steps toward clinical deployment of AI-based real-time quantitative detection of LrGG progression by RANO criteria. This quantitative assessment provides supplementary information to the radiologist and can be shared with the patient’s neuro-oncologist, neurosurgeon, and other clinical team members for consideration in treatment planning.