New ORCA method compresses 3D CT visual tokens for AI models

Researchers have developed ORCA (ORgan-Centroid Aggregation), a novel method for compressing visual tokens from 3D CT scans. This training-free approach merges adjacent tokens with organ guidance and incorporates centroid spatial encoding to preserve anatomical information. ORCA consistently outperforms existing compression methods at similar token budgets, significantly reducing the KV-cache size and processing time for downstream vision-language models. The code for ORCA has been released on Hugging Face. AI

IMPACT Enables more efficient processing of 3D medical imaging data by vision-language models, potentially improving diagnostic accuracy and report generation.

RANK_REASON The cluster contains a research paper detailing a new method for visual token compression in AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• 3d Ct
• CT-RATE
• Hugging Face
• Merlin
• ORCA
• ORgan-Centroid Aggregation
• vision-language model

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →