BT-CAP: a subcomponent-aware and anatomically constrained data augmentation framework for multi-modal brain tumor MRI...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Radiology

Background:

• Data scarcity and class imbalance hinder medical image analysis, especially for brain tumor MRI segmentation.
• Existing augmentation methods lack subcomponent control or require extensive generative model training.
• Underrepresentation of brain tumor subcomponents (enhancing tumor, non-enhancing tumor, cystic component, peritumoral edema) poses a significant challenge.

Purpose of the Study:

• To introduce the Brain Tumor Compositional Augmentation Pipeline (BT-CAP) for subcomponent-aware, anatomically constrained augmentation of multi-modal MRI.
• To address limitations of existing augmentation strategies in low-data settings for brain tumor segmentation.

Main Methods:

• BT-CAP decomposes and recomposes tumor subcomponents using targeted operations like scaling, B-spline deformation, and inpainting.
• The pipeline ensures consistent application across MRI modalities and segmentation masks, generating label-ready volumes.
• Augmentation diversity and anatomical plausibility were evaluated on 50 BraTS-PEDs 2025 cases, producing 250 augmented volumes.

Main Results:

• BT-CAP achieved high structural diversity (mean SSIM 0.956 ± 0.014) with realistic intensity variations.
• Augmented data maintained anatomical integrity, with masks confined within brain boundaries and no edema-tumor core overlap.
• Segmentation performance improved by 6%-7% for tumor subcomponents and 2%-3% for overall tumor segmentation.

Conclusions:

• BT-CAP offers a novel class of compositional augmentation for anatomically structured, label-ready data generation without generative models.
• The framework is scalable and applicable to multi-class segmentation tasks facing data scarcity and requiring structural fidelity.
• The BT-CAP framework is publicly available for research use.