NULLBUS: Multimodal Mixed-Supervision for Breast Ultrasound Segmentation via Nullable Global-Local Prompts
Published:Dec 25, 2025 05:00
•1 min read
•ArXiv Vision
Analysis
This paper introduces NullBUS, a novel framework addressing the challenge of limited metadata in breast ultrasound datasets for segmentation tasks. The core innovation lies in the use of "nullable prompts," which are learnable null embeddings with presence masks. This allows the model to effectively leverage both images with and without prompts, improving robustness and performance. The results, demonstrating state-of-the-art performance on a unified dataset, are promising. The approach of handling missing data with learnable null embeddings is a valuable contribution to the field of multimodal learning, particularly in medical imaging where data annotation can be inconsistent or incomplete. Further research could explore the applicability of NullBUS to other medical imaging modalities and segmentation tasks.
Key Takeaways
Reference
“We propose NullBUS, a multimodal mixed-supervision framework that learns from images with and without prompts in a single model.”