Low-Rank Adaptive Structural Priors for Generalizable Diabetic Retinopathy Grading
Yunxuan Wang 1 , Ray Yin 2 , Yumei Tan 3 , Hao Chen 4 , Haiying Xia 1 · 2025
Read Low-Rank Adaptive Structural Priors for Generalizable Diabetic Retinopathy Grading by Yunxuan Wang 1 , Ray Yin 2 , Yumei Tan 3 , Hao Chen 4 , Haiying Xia 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-19362.txt.
Introduction
Yunxuan Wang, Ray Yin, Yumei Tan, Hao Chen, and Haiying Xia present a 2025 paper on grading diabetic eye disease so the grade holds up at new hospitals. Deep learning grades diabetic retinopathy well on familiar data, but accuracy drops when scanners or populations shift. Most fixes ignore the vessel and lesion structures that grading depends on.
Their framework, LoASP, plugs into existing generalization methods and learns low-rank structural priors tuned to those vessel and lesion patterns. Experiments across eight datasets, in single-source and multi-source settings, confirmed the gains. Pictures of the learned priors line up with real vessel and lesion architecture. The authors argue structure-aware priors are a practical route to grading that travels.
Catalog record
| CREATOR | Yunxuan Wang 1 , Ray Yin 2 , Yumei Tan 3 , Hao Chen 4 , Haiying Xia 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-19362
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Low-Rank Adaptive Structural Priors for Generalizable Diabetic Retinopathy Grading Yunxuan Wang 1 , Ray Yin 2 , Yumei Tan 3 , Hao Chen 4 , Haiying Xia 1 Abstract Diabetic retinopathy (DR), a serious ocular complication of diabetes, is one of the primary causes of vision loss among retinal vascular diseases. Deep learning methods have been extensively applied in the grading of diabetic retinopathy (DR). However, their performance declines significantly when applied to data outside the training distribution due to domain shifts. Domain generalization (DG) has emerged as a solution to this challenge. However, most existing DG methods overlook lesion-specific features, resulting in insufficient accuracy. In this paper, we propose a novel approach that enhances existing DG methods by incorporating structural priors , inspired by the observation that DR grading is heavily dependent on vessel and lesion structures. We introduce Low-rank Adaptive Structural Priors (LoASP) , a plug-and-play framework designed for seamless integration with existing DG models. LoASP improves generalization by learning adaptive structural representations that are finely tuned to the complexities of DR diagnosis. Extensive experiments on eight diverse datasets validate its effectiveness in both single-source and multisource domain scenarios. Furthermore, visualizations reveal that the learned structural priors intuitively align with the intricate architecture of the vessels and lesions, providing compelling insights into their interpretability and diagnostic relevance. I Introduction Background. By 2040, it is estimated that around 600 million people worldwide will have diabetes, and one-third of these individuals will suffer from diabetic retinopathy (DR) [ 1 ] . DR, a severe complication of diabetes, progressively damages the retina’s delicate vascular networks. This damage can lead to changes such as angiogenesis and capillary regression, which may result in irreversible blindness if not timely detected and treated. Often, early-stage DR remains asymptomatic or results in minor vision disturbances, making early detection and accurate grading of DR crucial to preventing blindness [ 2 ] . Recent advancements in deep learning (DL) have significantly improved automated diagnosis in ophthalmology through medical image analysis. Current research on DR grading mainly focuses on the accuracy of detection of small lesions in the same dataset. These methods have achieved promising results in automated DR grading [ 3 , 4 ] , but neglect the data heterogeneity between datasets, i.e. , data shifts [ 5 , 6 ] , which results in the inability of the existing models to be directly applied to real-world clinical settings and poses a substantial obstacle in clinical practice [ 7 ] . Therefore, how to make models well adapted to out-of-distribution data is very crucial in clinical applications. Fig. 1: Overview of the LoASP framework. The LoASP framework comprises two key modules: the structural prior learning module (LoSP) and the adaptive projection module (LoAP) for effective prior integration. Our method is designed to excel in limited-data training scenarios from a few source domains (hospitals), aiming for superior performance and showcasing adaptability and robustness across diverse, unseen target domains. In the main figure, the black arrows represent the forward flow, while the blue dashed arrows illustrate the gradient backpropagation flow. The challenge of domain shifts has spurred the development of domain generalization (DG), a framework designed to enhance model robustness by enabling effective generalization across diverse domains without requiring target domain data during training [ 8 , 9 , 10 , 11 ] . In recent years, domain-invariant representations have gained attention, which employ regularization techniques to stabilize model behavior across domains [ 5 , 12 , 11 ] . Additionally, some studies have explored disentangling domain-specific [ 13 ] , domain-invariant [ 5 , 6 ] , or intra-class features [ 14 ] . Simultaneously, image augmentation techniques have been explored to simulate domain variability, using visual transformations or deliberate image degradation to enable models to tackle the diverse challenges posed by shifting domain characteristics [ 14 , 15 , 16 ] . Motivation. Existing DG methods largely focus on addressing domain shifts by emphasizing data characteristics, such as pixel intensity distributions (style shifts) or data imbalance. However, we argue that these methods neglect a crucial aspect: the diagnostic features that clinicians rely on in real-world settings to grade diabetic retinopathy (DR). These features, inherently invariant across domains, are central to accurate diagnosis and serve as a robust foundation for clinical decision-making. In this work, we go beyond conventional DG strategies by focusing on these domain-invariant diagnostic features specific to DR grading. The International Council of Ophthalmology (ICO) defines diabetic retinopathy (DR) stages through a five-grade system (Grade 0 to 4) based on fundus diagnosis, with a particular emphasis on blood vessel-related characteristics [ 17 ] : • Grade 0 (Normal Retina): Clear blood vessels with no visible lesions in the fundus images. • Grade 1 (Mild DR): Characterized by microaneurysms , appearing as small dark red dots in the retina. • Grade 2 (Moderate DR): Presence of microaneurysms, hemorrhages, and hard exudates . • Grade 3 (Severe DR): Includes abnormal blood vessels, extensive hemorrhages or blood clots, and soft exudates . • Grade 4 (Proliferative DR): Incorporates all prior lesions along with neovascularization , indicating advanced disease progression. Building upon this, we aim to utilize vessel-related characteristics as domain-invariant diagnostic features to address the domain shift problem. In this work, we propose Low-rank Adaptive Structural Priors (LoASP), comprising three primary components: the Low-rank Structural Prior (LoSP) module, the Low-rank Adaptive Projection (LoAP) module, and the Low-rank Design Mechanism. First, LoSP dynamically captures the structural information of the vessel. Then, LoAP efficiently fuses the imaging features with the structural prior via an adaptive B-spline projector. Finally, to further cope with the challenging nature of medical datasets, we incorporate a low-rank design into our module. Notably, our proposed LoASP is not only a promising solution for domain generalization (DG) but also a plug-and-play tool that enhances the performance of existing DG methods. In this work, we introduce a domain-invariant structural prior to develop a novel domain generalization module, LoASP, which consists of two key components: LoSP and LoAP. This innovative framework not only functions as an efficient standalone domain generalization module but also serves as a versatile, plug-and-play enhancement that can be seamlessly integrated with existing domain generalization techniques. Its adaptability makes it a promising solution for advancing diabetic retinopathy grading through adaptive domain generalization. Specifically: 1. We develop a vessel-related structural prior, termed LoSP, inspired by clinical observations in DR grading. This prior inherently captures domain-invariant characteristics, enabling robust and reliable analysis (Sec. III-A ). 2. To ensure efficient prior integration with current methods, we propose LoAP, a lightweight adapter featuring a learnable fusion projector. This projector adaptively projects the structural prior before fusing it with the original model output, enabling a more coherent and effective synergy (Sec. III-B ). 3. We verify our modules in several ways under domain generalization settings, first is the verify the module solely with model backbone, second is the to with the state-of-the-art domain generalization methods. We further done the detailed ablation into the design details of in-submodules and the parameters. We visualize the structural information and further solify our method in Sec. IV-D . II related work II-A Diabetic Retinopathy Analysis In recent years, deep learning has achieved remarkable advancements in DR