ReCoSeg: Residual-Guided Cross-Modal Diffusion for Efficient Brain Tumor Segmentation

SARA YAVARI, Rahul Pandya, Jacob Furst

Keywords: Brain Tumor Segmentation, Diffusion Models, Residual Maps, Semi-Supervised Learning.

Abstract:

Precise segmentation of brain tumors from MRI scans is important in clinical practice for effective diagnosis and proper treatment planning. Diffusion models have been highly effective in image generation and segmentation. In this work, we introduce ReCoSeg, a novel semi supervised framework that combines cross-modal diffusion based synthesis with residual guided segmentation to improve accuracy. First, a diffusion model synthesizes the T1ce modality from existing FLAIR, T1, and T2 MRI scans. The synthesized and the real image difference captured as residuals—highlights potential tumor regions. Residuals are then used as attention cues in a lightweight U-Net for segmentation, reducing the reliance on dense labels with an enhanced segmentation performance.

Submission Number: 79