ORBIT: Counterfactual Proposal Inference for Prompt-Free 3D Brain Tumor Segmentation

Jaelyn S. Liang

Keywords: Structured Inference, Counterfactual Generation, Zero-Shot Segmentation, 3D Medical Imaging, Label-Efficient Learning

TL;DR: ORBIT uses healthy counterfactual reconstruction and structured MAP mask selection to perform prompt-free zero-shot brain tumor segmentation in 3D MRI.

Abstract: Brain tumor MRI plays a central role in measuring disease extent and tracking change over time, yet fully automatic analysis of 3D studies remains limited by scarce voxel annotations and variable acquisition protocols. Supervised segmentation models depend on dense masks during training, while promptable foundation models often rely on boxes or clicks at test time. ORBIT addresses this setting with a prompt-free zero-shot pipeline for multi-sequence brain MRI. The method reconstructs a healthy counterfactual volume, forms tumor proposal fields from reconstruction residuals, and converts those proposals into a volumetric mask with self-tuned hysteresis over connected 3D support. The predicted mask also supports structured semantic evidence through WHO-aligned text descriptions and mask-derived radiology attributes. On BraTS 3D MRI, ORBIT reaches a macro Dice of 0.5776 under a reduced-input setting that uses no in-domain segmentation training and no interactive prompts. An ablation shows a 0.1315 absolute Dice improvement over naive residual thresholding, indicating that counterfactual proposal inference and structured mask selection are central to performance. The results highlight counterfactual proposal inference as a promising path for label-efficient 3D tumor analysis.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 211