Robust multimodal fusion network using adversarial learning for brain tumor grading

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Neuro-oncology

Background:

• Glioma grading relies on multimodal magnetic resonance imaging (MRI).
• Missing MRI modalities degrade glioma grading accuracy, impacting treatment and prognosis.
• There is a need for robust glioma grading models that can handle missing data.

Purpose of the Study:

• To develop and validate a robust brain tumor grading model capable of handling missing MRI modalities.
• To improve the accuracy and reliability of glioma grading when complete imaging data is unavailable.

Main Methods:

• A novel model utilizing adversarial learning to generate features for missing modalities.
• An attention-based fusion block to integrate features from available modalities.
• Nested five-fold cross-validation on the Brain Tumor Segmentation Challenge 2017 dataset (n=285).

Main Results:

• The proposed model achieved an average area under the curve of 87.76% across all missing-modality scenarios.
• Demonstrated superior performance in classifying high-grade gliomas (HGGs) and low-grade gliomas (LGGs) compared to competing methods.
• Activation maps confirmed clinically relevant focus on tumor enhancing/non-enhancing portions and edema.

Conclusions:

• The developed model provides robust glioma grading even with missing MRI modalities.
• This approach has potential to positively impact glioma care by providing reliable grading regardless of data completeness.