A fully automated MRI-based deep-learning algorithm for classifying germinomas and nongerminomatous germ cell tumors.

Abstract Intra-cranial germ cell tumors (iGCTs) are classified into two pathological subtypes Germinomas (GEs) and nongerminomatous germ cell tumors (NGGCTs). Accurate preoperative determination of iGCT subtypes is essential to guide clinical decision making and prognosis assessment. This study aims to develop a deep-learning algorithm to automatically segment iGCTs and classify subtypes using preoperative T2-weighted (T2W) images. Brain MR imaging and corresponding pathologic information were retrospectively obtained from 594 subjects including 269 GEs and 325 NGGCTs from Beijing Tiantan Hospital between January 1, 2008, and October 31, 2020, and subdivided into training (n=416) and test (n=178) sets for iGCTs segmentation, and classification using 3D nnU-Net based on T2W images. An internal validation cohort (including 56 GEs and 17 NGGCTs) and two external validation cohorts (Center 1 including 13 GEs and 12 NGGCTs；Center 2 including 17 GEs and 12 NGGCTs) were prospectively designed as an independent sets for testing the iGCT-net. Dice scores were computed to assess tumor segmentation. Accuracy and area under the curve (AUC) were calculated to assess the GEs and NGGCTs classification. Additionally, sensitivity analysis of different tumor location (suprasellar, pineal and basal ganglia regions) was performed to access the classification accuracy. The iGCT-net for tumor segmentation achieved a mean dice score of 0.80 in all cohorts. For tumor classification, the iGCT-net achieved the mean accuracy of 90.96% of all cohorts, which was superior to that using demographics and morphologic features (mean = 67.50%, p< 0.01). For different tumor location (suprasellar region, pineal region, and basal ganglia region), the iGCT-net showed an mean accuracy of 87.94%, 89.38% and 65.74% in test set, internal validation set and external validation cohorts. A fully automatic deep learning algorithm to segment iGCT and discriminate GEs and NGGCTs with high accuracy based on only T2W images from a large dataset is successfully developed.