Anatomy-Guided, Modality-Agnostic Segmentation of Neuroimaging Abnormalities.

Area of Science:

• Neuroimaging
• Machine Learning
• Radiology

Background:

• Magnetic resonance imaging (MRI) provides crucial brain views but often has variable sequence availability.
• This variability hinders radiological interpretation and limits machine learning model generalizability.
• Developing robust models for diverse MRI data is essential for accurate disease assessment.

Purpose of the Study:

• To propose an anatomy-guided, modality-agnostic framework for robust brain MRI abnormality detection.
• To enhance machine learning model performance under missing or variable imaging modality conditions.
• To improve the generalizability of abnormality detection in neuroimaging pipelines.

Main Methods:

• Developed an anatomy-guided, modality-agnostic framework for brain MRI analysis.
• Introduced Region ModalMix (RMM), an augmentation strategy integrating anatomical priors.
• Trained and evaluated the framework on BraTS 2020 and MU-Glioma-Post datasets.

Main Results:

• The RMM framework outperformed state-of-the-art methods on the BraTS 2020 dataset, reducing Hausdorff Distance (HD95) by 9.68 mm and improving Dice Similarity Coefficient (DSC) by 1.36%.
• On the MU-Glioma-Post dataset, RMM demonstrated strong out-of-distribution generalization, reducing HD95 by 18.24 mm and improving DSC by 9.54% in severe missing-modality scenarios.
• The framework proved effective in handling heterogeneous data availability.

Conclusions:

• The proposed framework ensures robust abnormality detection in brain MRI despite variable sequence availability.
• RMM enhances the generalizability of machine learning models in multimodal neuroimaging.
• This approach facilitates more reliable disease assessment in real-world clinical settings.