Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generalized Robust Fundus Photography-based Vision Loss Estimation for High Myopia

Zipei Yan, Zhile Liang, Zhengji Liu, Shuai Wang, Rachel Ka-Man Chun, Jizhou Li, Chea-su Kee, Dong Liang

TL;DR

High myopia elevates the risk of irreversible vision loss, yet VF assessment via perimetry is subjective and slow. The authors introduce a parameter-efficient Refinement-by-Denoising (RED) framework that refines features from pretrained vision models to produce robust, high-entropy representations for VF estimation, addressing domain gaps between fundus images and natural images. RED is trained with Stein's unbiased risk estimator (SURE) using Monte Carlo approximation (MC-SURE) to avoid ground-truth clean features, followed by a regression head that predicts VF. On two real-world HM datasets from distinct centers, RED achieves superior RMSE, MAE, and PCC in both in-distribution and out-of-distribution scenarios, indicating improved generalization and potential clinical utility in ophthalmic practice.

Abstract

High myopia significantly increases the risk of irreversible vision loss. Traditional perimetry-based visual field (VF) assessment provides systematic quantification of visual loss but it is subjective and time-consuming. Consequently, machine learning models utilizing fundus photographs to estimate VF have emerged as promising alternatives. However, due to the high variability and the limited availability of VF data, existing VF estimation models fail to generalize well, particularly when facing out-of-distribution data across diverse centers and populations. To tackle this challenge, we propose a novel, parameter-efficient framework to enhance the generalized robustness of VF estimation on both in- and out-of-distribution data. Specifically, we design a Refinement-by-Denoising (RED) module for feature refinement and adaptation from pretrained vision models, aiming to learn high-entropy feature representations and to mitigate the domain gap effectively and efficiently. Through independent validation on two distinct real-world datasets from separate centers, our method significantly outperforms existing approaches in RMSE, MAE and correlation coefficient for both internal and external validation. Our proposed framework benefits both in- and out-of-distribution VF estimation, offering significant clinical implications and potential utility in real-world ophthalmic practices.

Generalized Robust Fundus Photography-based Vision Loss Estimation for High Myopia

TL;DR

High myopia elevates the risk of irreversible vision loss, yet VF assessment via perimetry is subjective and slow. The authors introduce a parameter-efficient Refinement-by-Denoising (RED) framework that refines features from pretrained vision models to produce robust, high-entropy representations for VF estimation, addressing domain gaps between fundus images and natural images. RED is trained with Stein's unbiased risk estimator (SURE) using Monte Carlo approximation (MC-SURE) to avoid ground-truth clean features, followed by a regression head that predicts VF. On two real-world HM datasets from distinct centers, RED achieves superior RMSE, MAE, and PCC in both in-distribution and out-of-distribution scenarios, indicating improved generalization and potential clinical utility in ophthalmic practice.

Abstract

High myopia significantly increases the risk of irreversible vision loss. Traditional perimetry-based visual field (VF) assessment provides systematic quantification of visual loss but it is subjective and time-consuming. Consequently, machine learning models utilizing fundus photographs to estimate VF have emerged as promising alternatives. However, due to the high variability and the limited availability of VF data, existing VF estimation models fail to generalize well, particularly when facing out-of-distribution data across diverse centers and populations. To tackle this challenge, we propose a novel, parameter-efficient framework to enhance the generalized robustness of VF estimation on both in- and out-of-distribution data. Specifically, we design a Refinement-by-Denoising (RED) module for feature refinement and adaptation from pretrained vision models, aiming to learn high-entropy feature representations and to mitigate the domain gap effectively and efficiently. Through independent validation on two distinct real-world datasets from separate centers, our method significantly outperforms existing approaches in RMSE, MAE and correlation coefficient for both internal and external validation. Our proposed framework benefits both in- and out-of-distribution VF estimation, offering significant clinical implications and potential utility in real-world ophthalmic practices.
Paper Structure (14 sections, 2 theorems, 12 equations, 3 figures, 4 tables)

This paper contains 14 sections, 2 theorems, 12 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Proposed Method
  4. Overview
  5. Feature Refinement by Denoising (RED)
  6. Analysis of the role of noise
  7. Experiments
  8. Experimental Data
  9. Experimental Setup
  10. Experimental Results
  11. Ablation Study
  12. Conclusion
  13. Acknowledgements
  14. Disclosure of Interests

Key Result

theorem thmcountertheorem

(stein1981sure) The random variable $\text{\normalfont SURE}(h_{\boldsymbol{\theta}}(\boldsymbol{z}))$ is an unbiased estimator of $\text{\normalfont MSE}(h_{\boldsymbol{\theta}}(\boldsymbol{z}))$, that is, $\mathbb{E}_{\boldsymbol{n}}\{\text{\normalfont MSE}(h_{\boldsymbol{\theta}}(\boldsymbol{z}))

Figures (3)

  • Figure 1: Overview of the proposed method.
  • Figure 2: Visualization of predictions of a case in the Test set.
  • Figure 3: (a) L2 Norm of the weight of the regression layer (b) Entropy of feature space (c) MC-SURE loss during training, and (d) RMSE and MAE metrics corresponding to different $\lambda$.

Theorems & Definitions (2)

  • theorem thmcountertheorem
  • theorem thmcountertheorem