Robust deep learning for incomplete MRI sequences in glioma grading and IDH mutation status prediction: a large-scale multicenter study.
Authors
Affiliations (10)
Affiliations (10)
- Department of Radiology, School of Medicine, The Second Affiliated Hospital of South China University of Technology (Guangzhou First People's Hospital), Guangzhou, China.
- School of Medicine, South China University of Technology, Guangzhou, China.
- Department of MRI, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.
- School of Medical Equipment, Guangdong Food and Drug Vocational College, Guangzhou, China.
- Department of Radiology, Nanfang Hospital, Southern Medical University, Guangzhou, China.
- Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- Department of Radiology, Guangdong Sanjiu Brain Hospital, Guangdong, China.
- School of Biomedical Engineering, Southern Medical University, Guangzhou, China. [email protected].
- Department of Radiology, School of Medicine, The Second Affiliated Hospital of South China University of Technology (Guangzhou First People's Hospital), Guangzhou, China. [email protected].
- School of Medicine, South China University of Technology, Guangzhou, China. [email protected].
Abstract
Incomplete MRI sequences pose a significant challenge to the reliability of multiparametric MRI (mp-MRI) radiomics models. This study aimed to develop a robust and noninvasive approach for accurate glioma grading and isocitrate dehydrogenase (IDH) mutation status prediction using incomplete MRI data. Conventional MRI scans of 2170 glioma patients were retrospectively collected from five clinical institutions and a public dataset. Radiomic features extracted from each sequence, including incomplete ones, were processed using a robust incomplete sequence estimation network (RISEN). This model imputes missing features and learns latent fusion representations for glioma grading and IDH mutation status prediction. Model performance was evaluated by determining the optimal combination of mp-MRI sequences and simulating various clinical scenarios with different rates of missing data, using the area under the curve (AUC). The optimal sequence combination of T1WI, CE-T1WI, and T2-FLAIR achieved the highest performance for glioma grading (AUC: 0.8160, 0.9136, 0.8031) and IDH prediction (AUC: 0.8657, 0.8731, 0.7682) in internal and external validation datasets with complete data. In simulated incomplete-sequence scenarios, RISEN exhibited only moderate declines for glioma grading (mean AUC: 0.7942, 0.8955, 0.7759) and IDH prediction (mean AUC: 0.8454, 0.8478, 0.7414). Comparable robustness was observed in real-world missing-sequence data (AUC: 0.8910 and 0.8854, respectively). RISEN demonstrates robust and clinically applicable performance for glioma grading and IDH mutation prediction across multiple cohorts, even under commonly encountered incomplete mp-MRI conditions. Question Incomplete MRI sequences present a major challenge for radiomics-based glioma grading and IDH mutation status prediction in real-world clinical settings. Findings The proposed RISEN model imputes missing features using latent representations, improving model robustness and maintaining high predictive performance across various missing data scenarios. Clinical relevance RISEN offers a reliable solution for real-world glioma diagnosis when MRI sequences are incomplete, ensuring consistent clinical decision-making and personalized therapy planning across diverse imaging conditions.