Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Low-Rank Adaptive Structural Priors for Generalizable Diabetic Retinopathy Grading

Yunxuan Wang, Ray Yin, Yumei Tan, Hao Chen, Haiying Xia

TL;DR

DR grading is hampered by domain shifts across clinical datasets. The authors introduce LoASP, a plug-and-play framework comprising LoSP (structure learning) and LoAP (adaptive projection) that learns adaptive, domain-invariant vessel- and lesion-structure priors and fuses them with imaging features in a low-rank, efficient manner. Across eight DR datasets and multiple DG baselines, LoASP yields consistent improvements in AUC, ACC, and F1 for both DG and SDG settings, supported by visualizations that align learned priors with retinal vessels and lesions. The work provides a practical approach to robust DR grading under distribution shifts and points to the value of explicit structural priors for medical-domain generalization.

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 multi-source 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.

Low-Rank Adaptive Structural Priors for Generalizable Diabetic Retinopathy Grading

TL;DR

DR grading is hampered by domain shifts across clinical datasets. The authors introduce LoASP, a plug-and-play framework comprising LoSP (structure learning) and LoAP (adaptive projection) that learns adaptive, domain-invariant vessel- and lesion-structure priors and fuses them with imaging features in a low-rank, efficient manner. Across eight DR datasets and multiple DG baselines, LoASP yields consistent improvements in AUC, ACC, and F1 for both DG and SDG settings, supported by visualizations that align learned priors with retinal vessels and lesions. The work provides a practical approach to robust DR grading under distribution shifts and points to the value of explicit structural priors for medical-domain generalization.

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 multi-source 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.
Paper Structure (15 sections, 12 equations, 3 figures, 4 tables)

This paper contains 15 sections, 12 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. related work
  3. Diabetic Retinopathy Analysis
  4. Domain Generalization in Medical Imaging
  5. Low-rank Learning
  6. Methodology
  7. LoSP
  8. LoAP
  9. Complexity
  10. Experiments
  11. Experimental Design
  12. Model Generalization Experiments
  13. Ablation Study
  14. Visualization of Learned Priors
  15. Conclusion

Figures (3)

  • Figure 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.
  • Figure 2: Ablation Study on Parameter Selection. Red circles distinctly highlight and emphasize the chosen parameters for clear visualization.
  • Figure 3: Visualization of Learned Structural Priors. The first row presents the retinal imaging inputs, while the second row showcases the corresponding feature visualization results, illustrating the extracted structural priors.