PACE-MoE: Adaptive Domain Sampling and Domain-Specific LoRA for Medical Visual Grounding

Key points

• Medical visual grounding requires models to align clinical descriptions with spatial evidence across heterogeneous imaging domains.1
• PACE dynamically assesses the training value of each imaging domain from current-policy feedback and progressively adjusts domain coverage through domain admission, replay, and final all-domain balancing.1
• Experiments across seven medical imaging domains show that PACE-MoE consistently improves grounding performance over fixed-domain training, achieving a 3.64% relative improvement in mean box IoU over the Vanilla baseline.1
• These results establish coordinated data-parameter routing as an effective strategy for multi-domain medical visual grounding.1

Sources (1)

• [1]PACE-MoE: Adaptive Domain Sampling and Domain-Specific LoRA for Medical Visual Grounding
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 05:13 AM

  To address this issue, we propose PACE-MoE, a policy-guided framework that jointly performs adaptive domain sampling and domain-specific adaptation for medical visual grounding.

  Medical visual grounding requires models to align clinical descriptions with spatial evidence across heterogeneous imaging domains.

Extractive summary: sentences quoted from the sources.