Breast Ultrasound Diagnosis on BUSBRA in-domain (test)
77.8AUCUltrasoundAgents
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| UltrasoundAgentsBackbone=Qwen2.5-3B, Training Strategy=Three-stage RL (GRPO) and SFT, Inference Workflow=ROI -> attribute -> diagnosis evidence-chain2026.03 | 77.8 | 88.6 | 40 | 0.145 | |
| UltrasoundAgentsBackbone=Qwen2.5-3B, Training Strategy=Three-stage RL (GRPO) and SFT, Inference Workflow=ROI -> attribute -> diagnosis evidence-chain2026.03 | 72.3 | 81.3 | 62 | 0.3 | |
| Qwen2.5-3B-COT-SFTBackbone=Qwen2.5-3B, Training Strategy=SFT on template-based reasoning data, Inference Workflow=Intermediate attribute evidence2026.03 | 72.2 | 85.7 | 40 | 0.129 | |
| Qwen2.5-3B-COT-SFTBackbone=Qwen2.5-3B, Training Strategy=SFT on template-based reasoning data, Inference Workflow=Intermediate attribute evidence2026.03 | 66.8 | 72.2 | 56.3 | 0.258 | |
| Think-with-ImageBackbone=Qwen2.5-3B, Training Strategy=DeepEyes-style single-agent workflow, Inference Workflow=ROI localization -> crop-and-zoom -> re-analyze2026.03 | 50 | 69.3 | 4.8 | 0.003 | |
| Qwen2.5-3B-Zero-ShotBackbone=Qwen2.5-3B, Training Strategy=Zero-shot, Inference Workflow=Directly prompts on full image2026.03 | 50 | 74.3 | 20 | -0.053 | |
| Think-with-ImageBackbone=Qwen2.5-3B, Training Strategy=DeepEyes-style single-agent workflow, Inference Workflow=ROI localization -> crop-and-zoom -> re-analyze2026.03 | 50 | 74.3 | 31.4 | -0.001 | |
| Qwen2.5-3B-Zero-ShotBackbone=Qwen2.5-3B, Training Strategy=Zero-shot, Inference Workflow=Directly prompts on full image2026.03 | 45.8 | 58.8 | 9.1 | 0.003 |