Research map / Label-efficient and robust learning / Noisy labels
DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DiN, a conditional diffusion-based answer classifier for medical VQA trained with semantic noisy labels, plus a noisy-label benchmark built on VQA-RAD and PathVQA.
Problem
Medical VQA models assume clean labels, yet real Med-VQA annotations contain semantically structured label noise from human mislabeling, which existing noise-handling methods (designed for classification or natural VQA) handle poorly.
Approach
Recasts answer classification as a coarse-to-fine conditional diffusion process over answer distributions, conditioned on answer-aware features from a transformer-decoder Answer Condition Generator, and supervised by refined pseudo-labels from a robust focal-symmetric loss and dynamic answer adjustment (Noisy Label Refinement, used only in training).
Key results
- On VQA-RAD with 10% semantic noise, DiN reaches 74.24% overall accuracy versus 69.65% for the best baseline SNLC.
- On VQA-RAD with 20% semantic noise, DiN reaches 62.28% overall accuracy versus 58.88% for SNLC.
- On PathVQA with 20% semantic noise, DiN reaches 58.14% overall accuracy versus 56.72% for SNLC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiN instead of SNLC | Med-VQA · VQA-RAD 10%-Semantic Noise | overall accuracy: 69.65 -> 74.24 | Helped |
| Use DiN instead of SNLC | Med-VQA · VQA-RAD 20%-Semantic Noise | overall accuracy: 58.88 -> 62.28 | Helped |
| Use DiN instead of SNLC | Med-VQA · VQA-RAD 20%-Random Noise | overall accuracy: 62.03 -> 63.93 | Helped |
| Use DiN instead of SNLC | Med-VQA · PathVQA 10%-Semantic Noise | overall accuracy: 64.18 -> 66.67 | Helped |
| Use DiN instead of SNLC | Med-VQA · PathVQA 20%-Semantic Noise | overall accuracy: 56.72 -> 58.14 | Helped |
| Use DiN instead of SNLC | Med-VQA · PathVQA 20%-Random Noise | overall accuracy: 59.71 -> 60.84 | Helped |
| Add the Answer Diffuser (AD) (ablation) | DiN · VQA-RAD 10%-Semantic Noise | overall accuracy: 68.25 -> 71.15 | Helped |
| Add robust focal loss (RFL) (ablation) | DiN with AD · VQA-RAD 10%-Semantic Noise | overall accuracy: 71.15 -> 72.75 | Helped |
| Add answer adjustment (AA) (ablation) | DiN with AD · VQA-RAD 10%-Semantic Noise | overall accuracy: 71.15 -> 72.25 | Helped |
| Use RFL instead of CE, focal, or SCE loss (ablation) | DiN · VQA-RAD 10%-Semantic Noise | overall accuracy: 72.25; 73.10; 73.50 -> 74.24 | Helped |
| Replace MSE with KL loss for Ldif (ablation) | DiN · VQA-RAD 10%-Semantic Noise | overall accuracy: 74.24 -> 74.00 | No clear effect |
| Vary loss weight α from 0.5 to 0.1 or 1 (ablation) | DiN · VQA-RAD 10%-Semantic Noise | overall accuracy: 74.24; 73.75; 74.18 | No clear effect |
Surprising findings
- Diffusion steps T=50 performed best, far fewer than the 1000 typical in image generation, which the authors attribute to class-distribution rather than pixel-level generation.
- Under 20% noise, the cited SNLC method reached 62% on random noise but dropped to 58% on semantic noise, indicating semantic noise is harder.
Limitations
- Diffusion training may slightly increase training time.
- Evaluated only on two Med-VQA datasets (VQA-RAD and PathVQA) with simulated rather than real annotator noise.
Experimental setup
Datasets: VQA-RAD PathVQA
Benchmarks: VQA-RAD PathVQA
Metrics: accuracy (open, close, overall)
Baselines: MMBERT Q2ATransformer SimT CoDis SNLC Baseline (Swin+BERT+decoder)
Builds on: SNLC CARD Q2ATransformer MMBERT CoDis
Closest papers by meaning
- Alignment, Mining and Fusion: Representation Alignment with Hard Negative Mining and Selective Knowledge Fusion for Medical Visual Question Answering (CVPR 2025 · similarity 0.91)
- Falcon Medical Visual Question Answering (AAAI 2025 · similarity 0.90)
- Fine-grained Adaptive Visual Prompt for Generative Medical Visual Question Answering (AAAI 2025 · similarity 0.89)
- OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.89)
- Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning (ICML 2025 · similarity 0.88)
- Detecting Any instruction-to-answer interaction relationship:Universal Instruction-to-Answer Navigator for Med-VQA (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.