Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data
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.
Proposes GUES, a VAE-based generative unadversarial example generator trained with saliency-map supervision, to adapt diabetic retinopathy grading models in an online, model-agnostic setting without source data or model access.
Problem
Domain adaptation for diabetic retinopathy grading when the source model is absent (strictly controlled), target data arrives as a stream, and source data must remain private.
Approach
Reformulates iterative unadversarial perturbation optimization as a generative function of a latent input, instantiated with a VAE whose encoder predicts the latent and decoder generates per-image perturbations; saliency maps serve as pseudo-perturbation labels, justified by a theorem giving them as an upper bound on the latent input.
Key results
- GUES without training improves the source model on average over 12 transfer tasks by 4.5% ACC, 2.8% QWK and 3.7% AVG.
- GUES+SHOT-IM improves over SHOT-IM by over 3.0% ACC, 2.2% QWK and 2.6% AVG.
- With batch size 2, SHOT-IM drops about 16% ACC from batch 64 while GUES-combined methods show no evident decline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply GUES-generated unadversarial examples to the frozen source model | DeiT-base source model · 12 DR transfer tasks (Avg.) | ACC: 53.9 -> 58.4 (+4.5) | Helped |
| Apply GUES-generated unadversarial examples to the frozen source model | DeiT-base source model · 12 DR transfer tasks (Avg.) | QWK: 60.1 -> 62.9 (+2.8) | Helped |
| Preprocess online inputs with GUES before SHOT-IM | SHOT-IM · 12 DR transfer tasks (Avg.) | ACC: +3.0 | Helped |
| Preprocess online inputs with GUES before SHOT-IM | SHOT-IM · 12 DR transfer tasks (Avg.) | QWK: +2.2 | Helped |
| Preprocess online inputs with GUES before TENT | TENT · 12 DR transfer tasks (Avg.) | ACC: +1.7 | Helped |
| Preprocess online inputs with GUES before TENT | TENT · 12 DR transfer tasks (Avg.) | QWK: +4.1 | Helped |
| Remove both LKL and LMSE (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 51.0 | Hurt |
| Remove LMSE and retain LKL (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 54.7 | Hurt |
| Remove LKL and retain LMSE (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 56.9 | Hurt |
| Replace the VAE sampling component with a conventional autoencoder (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 53.8 | Hurt |
| Replace saliency-map supervision with a mix of saliency maps and original images (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 56.8 | Hurt |
| Replace saliency-map supervision with original images (ablation) | GUES · APTOS (Avg.) | ACC: 60.6 -> 56.5 | Hurt |
Surprising findings
- Replacing the input image with the saliency map (GUES w/ Sal) causes a large drop of about 15.4% average ACC.
Limitations
- Saliency-map supervision is argued to be effective for fundus images but may struggle on natural images with rich, task-irrelevant semantics.
- Evaluated only on DR grading fundus benchmarks with DeiT-base backbone.
Experimental setup
Datasets: APTOS DDR DeepDR Messidor-2
Benchmarks: APTOS DDR DeepDR Messidor-2
Metrics: ACC QWK AVG
Baselines: SHOT NRC CoWA PLUE TPDS SHOT-IM TENT SAR
Builds on: unadversarial examples SHOT TENT DeiT
Closest papers by meaning
- FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.89)
- Like an Ophthalmologist: Dynamic Selection Driven Multi-View Learning for Diabetic Retinopathy Grading (AAAI 2025 · similarity 0.86)
- Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging (AAAI 2025 · similarity 0.86)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.85)
- Samba: Severity-aware Recurrent Modeling for Cross-domain Medical Image Grading (NeurIPS 2024 · similarity 0.84)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.84)
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.