Molecular subgrouping of medulloblastoma based on few-shot learning of multitasking using conventional MR images: a retrospective multicenter study
Xi Chen, Zhen Fan, Kay Ka‐Wai Li, Guoqing Wu, Zhong Yang, Xin Gao, Yingchao Liu, Haibo Wu, Hong Chen, Qisheng Tang, Chen Liang, Yuanyuan Wang, Ying Mao, Ho‐Keung Ng, Zhifeng Shi, Jinhua Yu, Liangfu Zhou
Abstract
Abstract Background The determination of molecular subgroups—wingless (WNT), sonic hedgehog (SHH), Group 3, and Group 4—of medulloblastomas is very important for prognostication and risk-adaptive treatment strategies. Due to the rare disease characteristics of medulloblastoma, we designed a unique multitask framework for the few-shot scenario to achieve noninvasive molecular subgrouping with high accuracy. Methods We introduced a multitask technique based on mask regional convolutional neural network (Mask-RCNN). By effectively utilizing the comprehensive information including genotyping, tumor mask, and prognosis, multitask technique, on the one hand, realized multi-purpose modeling and simultaneously, on the other hand, promoted the accuracy of the molecular subgrouping. One hundred and thirteen medulloblastoma cases were collected from 4 hospitals during the 8-year period in the retrospective study, which were divided into 3-fold cross-validation cohorts (N = 74) from 2 hospitals and independent testing cohort (N = 39) from the other 2 hospitals. Comparative experiments of different auxiliary tasks were designed to illustrate the effect of multitasking in molecular subgrouping. Results Compared to the single-task framework, the multitask framework that combined 3 tasks increased the average accuracy of molecular subgrouping from 0.84 to 0.93 in cross-validation and from 0.79 to 0.85 in independent testing. The average area under the receiver operating characteristic curves (AUCs) of molecular subgrouping were 0.97 in cross-validation and 0.92 in independent testing. The average AUCs of prognostication also reached to 0.88 in cross-validation and 0.79 in independent testing. The tumor segmentation results achieved the Dice coefficient of 0.90 in both cohorts. Conclusions The multitask Mask-RCNN is an effective method for the molecular subgrouping and prognostication of medulloblastomas with high accuracy in few-shot learning.