Anatomy-guided, modality-agnostic segmentation of neuroimaging abnormalities.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Magnetic Resonance Imaging (MRI) provides crucial brain views but often has missing sequences in real-world data.
• Inconsistent multimodal MRI data challenges radiological interpretation and machine learning model generalizability.

Purpose of the Study:

• To develop an anatomy-guided, modality-agnostic framework for robust brain abnormality assessment despite variable MRI sequence availability.
• To introduce a novel data augmentation technique, Region ModalMix, to enhance model performance with incomplete multimodal inputs.

Main Methods:

• Proposed an anatomy-guided and modality-agnostic framework for brain MRI abnormality assessment.
• Introduced Region ModalMix, an augmentation strategy using anatomical priors for training robustness.
• Validated the framework on the BraTS 2020 dataset for brain tumor segmentation.

Main Results:

• The framework significantly outperformed baseline models under missing modality conditions.
• Achieved an average 9.68 mm reduction in 95th percentile Hausdorff Distance and 1.36% Dice Similarity Coefficient improvement.
• Demonstrated superior performance compared to state-of-the-art methods with incomplete data.

Conclusions:

• The proposed framework offers reliable abnormality detection in heterogeneous neuroimaging data settings.
• The method is model-agnostic, training-compatible, and broadly applicable to multi-modal pipelines.
• Enhances robustness and generalizability of AI models for brain MRI analysis.