Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty-Aware Ordinal Deep Learning for cross-Dataset Diabetic Retinopathy Grading

Ali El Bellaj, Aya Benradi, Salman El Youssoufi, Taha El Marzouki, Mohammed-Amine Cheddadi

TL;DR

This work tackles DR severity grading by addressing two key challenges: the ordinal nature of disease progression and the need for calibrated uncertainty under domain shift. It introduces a lesion-aware framework that combines a ConvNeXt backbone with a learnable lesion-query attention pooling mechanism and an evidential Dirichlet-based ordinal head, trained with an ordinal evidential loss and KL annealing. The approach achieves state-of-the-art-like accuracy and quadratic weighted kappa while providing meaningful uncertainty estimates, demonstrated through cross-dataset evaluation involving APTOS, Messidor-2, and EyePACS subsets. The findings suggest that integrating spatially focused lesion evidence, ordinal structure, and principled uncertainty is a promising direction for robust and clinically trustworthy DR screening systems.

Abstract

Diabetes mellitus is a chronic metabolic disorder characterized by persistent hyperglycemia due to insufficient insulin production or impaired insulin utilization. One of its most severe complications is diabetic retinopathy (DR), a progressive retinal disease caused by microvascular damage, leading to hemorrhages, exudates, and potential vision loss. Early and reliable detection of DR is therefore critical for preventing irreversible blindness. In this work, we propose an uncertainty-aware deep learning framework for automated DR severity grading that explicitly models the ordinal nature of disease progression. Our approach combines a convolutional backbone with lesion-query attention pooling and an evidential Dirichlet-based ordinal regression head, enabling both accurate severity prediction and principled estimation of predictive uncertainty. The model is trained using an ordinal evidential loss with annealed regularization to encourage calibrated confidence under domain shift. We evaluate the proposed method on a multi-domain training setup combining APTOS, Messidor-2, and a subset of EyePACS fundus datasets. Experimental results demonstrate strong cross-dataset generalization, achieving competitive classification accuracy and high quadratic weighted kappa on held-out test sets, while providing meaningful uncertainty estimates for low-confidence cases. These results suggest that ordinal evidential learning is a promising direction for robust and clinically reliable diabetic retinopathy grading.

Uncertainty-Aware Ordinal Deep Learning for cross-Dataset Diabetic Retinopathy Grading

TL;DR

This work tackles DR severity grading by addressing two key challenges: the ordinal nature of disease progression and the need for calibrated uncertainty under domain shift. It introduces a lesion-aware framework that combines a ConvNeXt backbone with a learnable lesion-query attention pooling mechanism and an evidential Dirichlet-based ordinal head, trained with an ordinal evidential loss and KL annealing. The approach achieves state-of-the-art-like accuracy and quadratic weighted kappa while providing meaningful uncertainty estimates, demonstrated through cross-dataset evaluation involving APTOS, Messidor-2, and EyePACS subsets. The findings suggest that integrating spatially focused lesion evidence, ordinal structure, and principled uncertainty is a promising direction for robust and clinically trustworthy DR screening systems.

Abstract

Diabetes mellitus is a chronic metabolic disorder characterized by persistent hyperglycemia due to insufficient insulin production or impaired insulin utilization. One of its most severe complications is diabetic retinopathy (DR), a progressive retinal disease caused by microvascular damage, leading to hemorrhages, exudates, and potential vision loss. Early and reliable detection of DR is therefore critical for preventing irreversible blindness. In this work, we propose an uncertainty-aware deep learning framework for automated DR severity grading that explicitly models the ordinal nature of disease progression. Our approach combines a convolutional backbone with lesion-query attention pooling and an evidential Dirichlet-based ordinal regression head, enabling both accurate severity prediction and principled estimation of predictive uncertainty. The model is trained using an ordinal evidential loss with annealed regularization to encourage calibrated confidence under domain shift. We evaluate the proposed method on a multi-domain training setup combining APTOS, Messidor-2, and a subset of EyePACS fundus datasets. Experimental results demonstrate strong cross-dataset generalization, achieving competitive classification accuracy and high quadratic weighted kappa on held-out test sets, while providing meaningful uncertainty estimates for low-confidence cases. These results suggest that ordinal evidential learning is a promising direction for robust and clinically reliable diabetic retinopathy grading.
Paper Structure (28 sections, 13 equations, 2 figures, 3 tables)

This paper contains 28 sections, 13 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Learning for Diabetic Retinopathy Grading
  4. Ordinal Learning in Medical Image Analysis
  5. Uncertainty Estimation and Evidential Deep Learning
  6. Cross-Dataset Generalization in Fundus Imaging
  7. Positioning of This Work
  8. Methodology
  9. Data Preprocessing and Quality Control
  10. Quality Filtration
  11. Normalization and Augmentation Strategy
  12. Backbone and Hierarchical Feature Extraction
  13. Lesion-Query Attention Pooling
  14. Transformer Decoder Architecture
  15. Interpretability and Feature Evidence
  16. ...and 13 more sections

Figures (2)

  • Figure 1: Overview of the proposed lesion-aware ordinal evidential architecture. A pretrained ConvNeXt-Base backbone extracts spatial feature maps that are projected into token embeddings and augmented with positional signals. Learnable lesion queries perform cross-attention pooling to selectively aggregate discriminative retinal regions, producing global representation. An ordinal evidential head then estimates Dirichlet parameters to jointly predict class probabilities and epistemic uncertainty.
  • Figure 2: Evolution of lesion-query specialization during training. (a) Early epoch activations show a lack of differentiation, whereas (b) late epoch activations reveal the model's transition toward independent feature extraction for more robust grading.