Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MAST-Pro: Dynamic Mixture-of-Experts for Adaptive Segmentation of Pan-Tumors with Knowledge-Driven Prompts

Runqi Meng, Sifan Song, Pengfei Jin, Yujin Oh, Lin Teng, Yulin Wang, Yiqun Sun, Ling Chen, Xiang Li, Quanzheng Li, Ning Guo, Dinggang Shen

TL;DR

MAST-Pro tackles pan-tumor segmentation by addressing tumor heterogeneity and dataset imbalance through a dynamic Mixture-of-Experts (D-MoE) framework augmented with knowledge-driven prompts. Text prompts from an LLM and anatomical prompts from TotalSegmentator provide domain priors, while D-MoE adaptively balances generic and tumor-specific features via task-dependent routing and cross-attention with image features. A Parameter-Efficient Fine-Tuning (PEFT) strategy updates only a subset of experts to enable scalable adaptation. Across eight multi-anatomical datasets, MAST-Pro achieves up to $5.20\%$ DSC improvement over SOTA while reducing trainable parameters by $91.04\%$, demonstrating both high accuracy and substantial efficiency for clinical deployment.

Abstract

Accurate tumor segmentation is crucial for cancer diagnosis and treatment. While foundation models have advanced general-purpose segmentation, existing methods still struggle with: (1) limited incorporation of medical priors, (2) imbalance between generic and tumor-specific features, and (3) high computational costs for clinical adaptation. To address these challenges, we propose MAST-Pro (Mixture-of-experts for Adaptive Segmentation of pan-Tumors with knowledge-driven Prompts), a novel framework that integrates dynamic Mixture-of-Experts (D-MoE) and knowledge-driven prompts for pan-tumor segmentation. Specifically, text and anatomical prompts provide domain-specific priors, guiding tumor representation learning, while D-MoE dynamically selects experts to balance generic and tumor-specific feature learning, improving segmentation accuracy across diverse tumor types. To enhance efficiency, we employ Parameter-Efficient Fine-Tuning (PEFT), optimizing MAST-Pro with significantly reduced computational overhead. Experiments on multi-anatomical tumor datasets demonstrate that MAST-Pro outperforms state-of-the-art approaches, achieving up to a 5.20% improvement in average DSC while reducing trainable parameters by 91.04%, without compromising accuracy.

MAST-Pro: Dynamic Mixture-of-Experts for Adaptive Segmentation of Pan-Tumors with Knowledge-Driven Prompts

TL;DR

MAST-Pro tackles pan-tumor segmentation by addressing tumor heterogeneity and dataset imbalance through a dynamic Mixture-of-Experts (D-MoE) framework augmented with knowledge-driven prompts. Text prompts from an LLM and anatomical prompts from TotalSegmentator provide domain priors, while D-MoE adaptively balances generic and tumor-specific features via task-dependent routing and cross-attention with image features. A Parameter-Efficient Fine-Tuning (PEFT) strategy updates only a subset of experts to enable scalable adaptation. Across eight multi-anatomical datasets, MAST-Pro achieves up to DSC improvement over SOTA while reducing trainable parameters by , demonstrating both high accuracy and substantial efficiency for clinical deployment.

Abstract

Accurate tumor segmentation is crucial for cancer diagnosis and treatment. While foundation models have advanced general-purpose segmentation, existing methods still struggle with: (1) limited incorporation of medical priors, (2) imbalance between generic and tumor-specific features, and (3) high computational costs for clinical adaptation. To address these challenges, we propose MAST-Pro (Mixture-of-experts for Adaptive Segmentation of pan-Tumors with knowledge-driven Prompts), a novel framework that integrates dynamic Mixture-of-Experts (D-MoE) and knowledge-driven prompts for pan-tumor segmentation. Specifically, text and anatomical prompts provide domain-specific priors, guiding tumor representation learning, while D-MoE dynamically selects experts to balance generic and tumor-specific feature learning, improving segmentation accuracy across diverse tumor types. To enhance efficiency, we employ Parameter-Efficient Fine-Tuning (PEFT), optimizing MAST-Pro with significantly reduced computational overhead. Experiments on multi-anatomical tumor datasets demonstrate that MAST-Pro outperforms state-of-the-art approaches, achieving up to a 5.20% improvement in average DSC while reducing trainable parameters by 91.04%, without compromising accuracy.

Paper Structure

This paper contains 11 sections, 8 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Prompt Embedding
  4. Pan-tumor Adaptive Mixture-of-Experts
  5. Parameter-Efficient Fine-Tuning (PEFT) strategy
  6. Experiment
  7. Dataset
  8. Implementation Details
  9. Comparison with State-of-the-art Methods
  10. Ablation Study
  11. Conclusion

Figures (2)

  • Figure 1: An overview of the proposed MAST-Pro model for pan-tumor segmentation, with the text and anatomy prompts served as specific priors to enhance tumor representation learning. Multi-anatomical radiology images are processed through D-MoE-enhanced SwinUNETR, where task-dependent routers dynamically select experts to balance generic and tumor-specific feature learning.
  • Figure 2: Qualitative visualizations comparing the proposed MAST-Pro model with other prompt-driven methods for multi-tumor segmentation. The first column shows the original CT scans, while the second column presents the ground-truth segmentations. The segmentation results in rows one to four correspond to liver tumors, kidney tumors, lung tumors, and colon tumors, respectively.