MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models

2026-03-12

AI Summary

Existing 3D medical vision-language models suffer from low computational efficiency due to substantial anatomical redundancy introduced by naively concatenating consecutive 2D slices and the use of fixed token pruning ratios that fail to account for varying information density across slices. To address this, this work proposes a training-free, model-agnostic hierarchical token pruning framework. The approach first eliminates inter-slice redundancy through anchor-based filtering and then dynamically selects the most informative tokens based on cumulative attention weights, enabling adaptive compression. By moving beyond rigid, fixed-ratio pruning, the method achieves remarkable efficiency: on three 3D medical benchmarks and across multiple medical vision-language models, it retains less than 5% of the original visual tokens while preserving or even surpassing baseline performance, substantially accelerating inference.

Technology Category

• Computer Vision: 3D Computer Vision
• Natural Language Processing: Summarization
• Machine Learning: Learning on the Edge & Model Compression

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

While specialized Medical Vision-Language Models (VLMs) have achieved remarkable success in interpreting 2D and 3D medical modalities, their deployment for 3D volumetric data remains constrained by significant computational inefficiencies. Current architectures typically suffer from massive anatomical redundancy due to the direct concatenation of consecutive 2D slices and lack the flexibility to handle heterogeneous information densities across different slices using fixed pruning ratios. To address these challenges, we propose MedPruner, a training-free and model-agnostic hierarchical token pruning framework specifically designed for efficient 3D medical image understanding. MedPruner introduces a two-stage mechanism: an Inter-slice Anchor-based Filtering module to eliminate slice-level temporal redundancy, followed by a Dynamic Information Nucleus Selection strategy that achieves adaptive token-level compression by quantifying cumulative attention weights. Extensive experiments on three 3D medical benchmarks and across three diverse medical VLMs reveal massive token redundancy in existing architectures. Notably, MedPruner enables models such as MedGemma to maintain or even exceed their original performance while retaining fewer than 5% of visual tokens, thereby drastically reducing computational overhead and validating the necessity of dynamic token selection for practical clinical deployment. Our code will be released.

Problem

• 3D medical image understanding
• computational inefficiency
• anatomical redundancy
• heterogeneous information density
• token pruning

Innovation

• token pruning
• 3D medical imaging
• vision-language models
• training-free
• dynamic attention