Guo Xinji, Lai Mingyao, Zhang Jie, Wang Lichao, Huang Haiyang, Huang Lijun, Li Hainan, Cai Linbo, Liang Jiuxing
Institute for Brain Research and Rehabilitation, South China Normal University, Guangzhou 510631, China.
Department of Oncology, Guangdong Sanjiu Brain Hospital, Guangzhou 510510, China.
Comput Methods Programs Biomed. 2025 Oct;270:108936. doi: 10.1016/j.cmpb.2025.108936. Epub 2025 Jul 2.
Glioblastoma (GBM) is an aggressive and highly malignant brain tumor, with isocitrate dehydrogenase (IDH) wild-type GBM being particularly notable for its poor prognosis. We developed the IDH wild-type Glioblastoma Integrated Survival Model (iGISM), a multimodal framework combining magnetic resonance imaging (MRI) and whole slide imaging (WSI) data for survival prediction in this subtype.
This retrospective study enrolled 180 patients with IDH wild-type GBM. The dataset included 180 H&E-stained WSIs and 720 preoperative MRI scans (T1, T1C, T2, and FLAIR sequences), and was randomly split into training (N = 144) and validation (N = 36) cohorts. The iGISM comprises of two pathways: an MRI pathway and a WSI pathway. In the MRI pathway, a ResNet18 backbone was employed, augmented with coordinate attention module and feature aggregation module. In the WSI pathway, we introduced a novel patch sampling strategy that combines image information entropy and nuclear density. An MRI-guided Co-Attention module was designed to facilitate cross-modality interactions. The extracted MRI and WSI features were integrated to generate the final multimodal risk score.
The iGISM achieved a concordance index (C-index) of 0.791 in the validation cohort, outperforming the MRI-only model (C-index: 0.713) and the WSI-only models (C-index range: 0.530-0.591). It also yielded the lowest integrated Brier score (IBS, 0.129), with time-dependent AUCs at 1, 2, and 3 years (0.821, 0.686, and 0.860, respectively). By comparison, the proposed WSI sampling strategy achieved the highest 17.7 % improvement in predictive performance over established methods.
The iGISM advances prognostic accuracy for IDH wild-type GBM by integrating MRI and WSI data, providing a valuable tool for guiding personalized treatment strategies in this patient population.
胶质母细胞瘤（GBM）是一种侵袭性很强的高度恶性脑肿瘤，其中异柠檬酸脱氢酶（IDH）野生型GBM的预后尤其差。我们开发了IDH野生型胶质母细胞瘤综合生存模型（iGISM），这是一个结合磁共振成像（MRI）和全切片成像（WSI）数据的多模态框架，用于预测该亚型的生存情况。
这项回顾性研究纳入了180例IDH野生型GBM患者。数据集包括180张苏木精-伊红染色的全切片图像和720次术前MRI扫描（T1、T1C、T2和FLAIR序列），并随机分为训练队列（N = 144）和验证队列（N = 36）。iGISM由两条路径组成：一条MRI路径和一条WSI路径。在MRI路径中，采用了ResNet18主干，并辅以坐标注意力模块和特征聚合模块。在WSI路径中，我们引入了一种结合图像信息熵和核密度的新型补丁采样策略。设计了一个MRI引导的协同注意力模块，以促进跨模态交互。提取的MRI和WSI特征被整合以生成最终的多模态风险评分。
iGISM在验证队列中的一致性指数（C指数）为0.791，优于仅使用MRI的模型（C指数：0.713）和仅使用WSI的模型（C指数范围：0.530 - 0.591）。它还产生了最低的综合Brier评分（IBS，0.129），1年、2年和3年的时间依赖性AUC分别为0.821、0.686和0.860。相比之下，所提出的WSI采样策略在预测性能上比现有方法提高了17.7%，是提高幅度最大的。
iGISM通过整合MRI和WSI数据提高了IDH野生型GBM的预后准确性，为指导该患者群体的个性化治疗策略提供了一个有价值的工具。
