WaveDE: Wavelet-Routed Dual-Expert Network for Fundus Photograph Enhancement
Haitao Nie*, Dongjie Wu*, and Bin Xie✉. Accepted at MICCAI 2026. Equal contribution.
View codeAI for Science · Medical Imaging · Trustworthy AI
Researching trustworthy AI for medical imaging.
I am Dongjie Wu, a Computer Science undergraduate at Central South University. My work spans medical image restoration, generative models, foundation-model data curation, and robust multimodal AI evaluation.
B.E. candidate · Central South University · 2027
Research at a glance
+4.13dB
PSNR over MDA-Net, alongside consistent gains across structural and perceptual metrics.
24
public datasets integrated into a reproducible degradation, training, and benchmarking pipeline.
~20K
retinal cases curated for the Global RETFound Consortium’s Hunan ophthalmic cohort.
Selected output
Publications and intellectual property
Research output spanning frequency-aware restoration, controllable diffusion, and degradation-aware enhancement for medical imaging.
Controllable Region-Aware Diffusion for High-Fidelity Fundus Enhancement
Haitao Nie*, Dongjie Wu*, Bin Xie✉, et al. Manuscript in preparation; target journal: Nature Medicine Intelligence.
Degradation-Aware Fundus Image Quality Enhancement Method, System, Device, and Medium
Bin Xie, Yiran Dai, Dongjie Wu, and Haitao Nie. China patent application CN 202610656256.0.
Research experience
From medical image restoration to robust multimodal AI
Four projects across image restoration, generative modeling, foundation-model data curation, and medical AI safety.
Controllable Region-Aware Diffusion for High-Fidelity Fundus Enhancement
- Proposed a controllable region-aware diffusion framework that restores clinician-specified regions while preserving non-selected areas, addressing precise enhancement with minimal hallucination.
- Achieved +4.13 dB PSNR over MDA-Net with consistent gains on SSIM, LPIPS, VIF, and Haralick; outperformed 12 recent baselines across in-domain and OOD settings.
- Built an end-to-end pipeline for compound degradation modeling, training, and benchmarking across 24 public datasets, supporting a manuscript currently in preparation.
WaveDE: Wavelet-Routed Dual-Expert Network
- Designed a wavelet-routed dual-expert architecture with frequency-aware routing to resolve the optimization conflict between high-frequency detail and low-frequency color and illumination.
- Improved detail recovery through HF-injected VAE decoding and enforced structural fidelity through anatomical consistency regularization.
- Reached state-of-the-art PSNR and SSIM among 12 methods from leading venues, with downstream gains that led to a MICCAI 2026 publication.
Global RETFound: Hunan Ophthalmic Data Curation
- Built a model-ready Hunan ophthalmic cohort from the Second Xiangya Hospital for the Global RETFound Consortium.
- Curated approximately 20,000 retinal cases through acquisition, cleaning, metadata standardization, image–text alignment, and quality control.
- Covered clinically relevant variables including demographics, camera type, diabetic retinopathy, AMD, glaucoma, hypertension, and diabetes.
Medical VLM Benchmarking for Robustness and Safety
- Developed a multidimensional Med-VLM benchmark spanning accuracy, robustness, fairness, and risk aversion under clinical uncertainty and distribution shifts.
- Built LLM-assisted label-verification and safety-evaluation workflows for medical descriptors and Chinese LLM jailbreak benchmarking.
Education
Central South University
Bachelor of Engineering in Computer Science and Technology · Double First-Class Initiative · Project 985
88.21
GPA out of 100
2×
Annual Scholarship
95
Machine Learning
- Digital Image Processing94 / 100
- Machine Learning95 / 100
- Artificial Intelligence89 / 100
Awards & certifications
- National Finalist
- CCF-CSP · 270 points
- National Third Prize
- Third Prize · AIGC Innovation
- Annual Scholarship
Connect
Let’s discuss AI for medical imaging.
I welcome conversations about AI for Science, trustworthy medical imaging, research collaboration, and graduate study.