Document Type : Applied Article

Author

• Farkhondeh Kiaee

Department of Electrical Engineering, Technical and Vocational University (TVU), Tehran, Iran.

Abstract

Brain tumor segmentation from multi-modal MRI remains challenging due to limited annotations, severe class imbalance, and substantial tumor heterogeneity. We propose a generation–distillation framework for foundation-aligned brain tumor segmentation under limited supervision. The framework combines (i) a prior-guided synthetic data generation engine, which produces anatomically plausible glioma MRI volumes and tumor masks through anatomically constrained tumor-mask diffusion, tumor growth simulation, and pathology-aware rendering, with (ii) DINOv3-based representation distillation, which transfers transferable visual representations from a frozen foundation teacher to a lightweight hierarchical 3D student network. The generated data are used to enrich representation pretraining, while a limited set of real annotated MRI volumes is used for downstream adaptation. Multi-scale, structural, domain, and cross-modal consistency objectives further promote anatomically coherent and modality-aware representations. Experiments on BraTS2020 demonstrate that the proposed framework produces diverse and anatomically plausible synthetic tumors and consistently improves segmentation performance over strong pretrained and state-of-the-art semi-supervised methods under limited supervision. External evaluation on the UCSF-PDGM and UPENN-GBM datasets further demonstrates improved cross-dataset generalization to independent clinical cohorts.

Keywords

• Brain Tumor Segmentation
• Foundation Models
• Knowledge Distillation
• Multi-modal Magnetic Resonance Imaging (MRI)
• Synthetic Data Generation

Main Subjects

• H.5. Image Processing and Computer Vision