Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ordinal Label-Distribution Learning with Constrained Asymmetric Priors for Imbalanced Retinal Grading

Nagur Shareef Shaik, Teja Krishna Cherukuri, Adnan Masood, Ehsan Adeli, Dong Hye Ye

TL;DR

This work tackles the challenge of ordinal, imbalanced DR grading by introducing CAP-WAE, a Wasserstein autoencoder variant with a constrained asymmetric prior and geometry-aware latent regularization. The model combines an AGGD-based latent prior, three specialized heads (classification, asymmetric Gaussian soft labels, and ordinal regression), and MAOC to produce grade-ordered, well-separated latent manifolds. An adaptive, uncertainty-weighted objective fuses reconstruction, aggregate prior alignment, and ordinal supervision, achieving state-of-the-art results on four public DR benchmarks with improved QWK, accuracy, and macro-F1, as well as qualitative evidence of structured latent geometry. The approach offers a robust template for ordinal, imbalanced medical grading and holds promise for extending to other diseases and longitudinal progression modeling.

Abstract

Diabetic retinopathy grading is inherently ordinal and long-tailed, with minority stages being scarce, heterogeneous, and clinically critical to detect accurately. Conventional methods often rely on isotropic Gaussian priors and symmetric loss functions, misaligning latent representations with the task's asymmetric nature. We propose the Constrained Asymmetric Prior Wasserstein Autoencoder (CAP-WAE), a novel framework that addresses these challenges through three key innovations. Our approach employs a Wasserstein Autoencoder (WAE) that aligns its aggregate posterior with a asymmetric prior, preserving the heavy-tailed and skewed structure of minority classes. The latent space is further structured by a Margin-Aware Orthogonality and Compactness (MAOC) loss to ensure grade-ordered separability. At the supervision level, we introduce a direction-aware ordinal loss, where a lightweight head predicts asymmetric dispersions to generate soft labels that reflect clinical priorities by penalizing under-grading more severely. Stabilized by an adaptive multi-task weighting scheme, our end-to-end model requires minimal tuning. Across public DR benchmarks, CAP-WAE consistently achieves state-of-the-art Quadratic Weighted Kappa, accuracy, and macro-F1, surpassing both ordinal classification and latent generative baselines. t-SNE visualizations further reveal that our method reshapes the latent manifold into compact, grade-ordered clusters with reduced overlap.

Ordinal Label-Distribution Learning with Constrained Asymmetric Priors for Imbalanced Retinal Grading

TL;DR

This work tackles the challenge of ordinal, imbalanced DR grading by introducing CAP-WAE, a Wasserstein autoencoder variant with a constrained asymmetric prior and geometry-aware latent regularization. The model combines an AGGD-based latent prior, three specialized heads (classification, asymmetric Gaussian soft labels, and ordinal regression), and MAOC to produce grade-ordered, well-separated latent manifolds. An adaptive, uncertainty-weighted objective fuses reconstruction, aggregate prior alignment, and ordinal supervision, achieving state-of-the-art results on four public DR benchmarks with improved QWK, accuracy, and macro-F1, as well as qualitative evidence of structured latent geometry. The approach offers a robust template for ordinal, imbalanced medical grading and holds promise for extending to other diseases and longitudinal progression modeling.

Abstract

Diabetic retinopathy grading is inherently ordinal and long-tailed, with minority stages being scarce, heterogeneous, and clinically critical to detect accurately. Conventional methods often rely on isotropic Gaussian priors and symmetric loss functions, misaligning latent representations with the task's asymmetric nature. We propose the Constrained Asymmetric Prior Wasserstein Autoencoder (CAP-WAE), a novel framework that addresses these challenges through three key innovations. Our approach employs a Wasserstein Autoencoder (WAE) that aligns its aggregate posterior with a asymmetric prior, preserving the heavy-tailed and skewed structure of minority classes. The latent space is further structured by a Margin-Aware Orthogonality and Compactness (MAOC) loss to ensure grade-ordered separability. At the supervision level, we introduce a direction-aware ordinal loss, where a lightweight head predicts asymmetric dispersions to generate soft labels that reflect clinical priorities by penalizing under-grading more severely. Stabilized by an adaptive multi-task weighting scheme, our end-to-end model requires minimal tuning. Across public DR benchmarks, CAP-WAE consistently achieves state-of-the-art Quadratic Weighted Kappa, accuracy, and macro-F1, surpassing both ordinal classification and latent generative baselines. t-SNE visualizations further reveal that our method reshapes the latent manifold into compact, grade-ordered clusters with reduced overlap.

Paper Structure

This paper contains 33 sections, 14 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Notation.
  4. Ordinal Label-Distribution Learning (OLDL)
  5. Order-aware divergences.
  6. Asymmetric Label Distributions in Medical Grading
  7. Latent Distribution Learning
  8. VAEs.
  9. WAEs and aggregate alignment.
  10. Asymmetric Generalized Gaussian (AGGD): Conceptual Background
  11. Methodology
  12. Architecture
  13. Encoder-Decoder.
  14. Prediction Heads.
  15. Constrained Asymmetric Prior
  16. ...and 18 more sections

Figures (2)

  • Figure 1: (a) Direction-aware ordinal label distributions allocate asymmetric probability mass around the reference grade. (b) Latent manifolds: a spherical Gaussian prior (left) contracts minority modes and overlaps grades, whereas a constrained asymmetric prior (right) preserves skew/tails and yields grade-ordered separability.
  • Figure 2: t-SNE latent-space evolution across ablation. Left: VAE–KL baseline shows overlapping clusters and weak ordinal structure. Middle: WAE–MMD improves separation but retains cross-class mixing. Right: Full CAP–WAE (ours) exhibits compact, well-ordered bands along disease severity with clearer orthogonal separation.