Robust deep learning for incomplete MRI sequences in glioma grading and IDH mutation status prediction: a large-scale...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Incomplete multiparametric MRI (mp-MRI) sequences hinder the reliability of radiomics models for glioma grading and IDH mutation status prediction.
• Developing robust models that can handle missing data is crucial for clinical application.

Purpose of the Study:

• To develop a robust and noninvasive approach for accurate glioma grading and IDH mutation status prediction using incomplete MRI data.
• To evaluate the performance of the proposed model across various clinical scenarios with missing MRI sequences.

Main Methods:

• Retrospective collection of conventional MRI scans from 2170 glioma patients across multiple institutions and a public dataset.
• Utilizing a robust incomplete sequence estimation network (RISEN) to impute missing radiomic features and learn latent representations.
• Evaluating model performance using the area under the curve (AUC) in complete, simulated incomplete, and real-world missing-sequence scenarios.

Main Results:

• The optimal mp-MRI sequence combination (T1WI, CE-T1WI, T2-FLAIR) achieved high performance in complete datasets.
• RISEN demonstrated only moderate performance declines in simulated incomplete-sequence scenarios for both glioma grading and IDH prediction.
• The model showed comparable robustness in real-world datasets with missing sequences.

Conclusions:

• RISEN provides a robust and clinically applicable solution for glioma grading and IDH mutation prediction, even with incomplete mp-MRI data.
• The model's ability to handle missing sequences ensures consistent clinical decision-making and personalized therapy planning.