Beyond proof-of-concept: Validating robust automated diffuse lower-grade glioma segmentation for clinical applications in longitudinal follow-up

Abstract Background Monitoring diffuse lower-grade glioma (DLGG) evolution on MRI is challenging due to their infiltrative nature and post-surgical deformations. The RANO criteria recommend assessing the 2D tumor size, but volume analysis remains the gold standard for calculating growth rate. Manual tumor segmentation, however, is time-consuming, limiting its use in clinical practice. Automated segmentation tools like nnU-Net are promising but require clinical validation. Methods We used 1971 MRI exams from 207 DLGG patients to train and validate an in house-developed nnU-Net al.orithm. The dataset included scans from various MRI systems, with 2D and 3D FLAIR and T1-weighted acquisitions, that were divided into derivation (N = 1771) and validation (N = 200) sets with matching 2D and 3D FLAIR ratios. The algorithm’s automated segmentations (AS) were compared with manual segmentations (MS) by expert neuroradiologists using the dice similarity coefficient (DSC) and Intersection over Union (IoU). Tumor volume and mean tumor diameter (MTD) were compared using Lin’s concordance correlation coefficient (CCC) and Bland-Altman tests. Results The median nnU-Net DSC and IoU were 0.93 and 0.86, respectively. In the validation cohort, 64% of exams had excellent (≥0.9), 31% good ([0.7-0.9]), 3.5% unsatisfactory ([0.5-0.7]), and 1.5% poor DSC (<0.5). Higher DSC correlated with larger tumor volumes (P < .001). Tumor volume and MTD showed near-perfect concordance between AS and MS (CCC: 0.991 and 0.989, respectively). Conclusions Our automated nnU-Net segmentation tool demonstrates high accuracy and potential for clinical integration, enhancing DLGG monitoring and management.