OUCopula: Bi-Channel Multi-Label Copula-Enhanced Adapter-Based CNN for Myopia Screening Based on OU-UWF Images

Yang Li, Qiuyi Huang, Chong Zhong, Danjuan Yang, Meiyan Li, Alan H. Welsh, Aiyi Liu, Bo Fu, Catherine C. Liu, Xingtao Zhou

Abstract

Myopia is a significant global health concern, particularly in Asia. This paper presents OUCopula, a deep learning framework designed for myopia screening using Ultra-Widefield (UWF) fundus images of both eyes (Oculus Uterque). By leveraging a bi-channel CNN with residual adapters and a copula-likelihood loss function, the model effectively captures interocular correlations and the statistical dependency between clinical markers like Spherical Equivalent (SE) and Axial Length (AL). Experimental results demonstrate that OUCopula significantly outperforms traditional single-eye models and standard CNN architectures, achieving a 7.18% improvement in Mean Square Error (MSE).

Introduction

Myopia (nearsightedness) has become a global health crisis, with early screening being essential to prevent severe complications such as retinal detachment. Ultra-Widefield (UWF) fundus imaging offers a comprehensive 200° view of the retina, but its interpretation requires specialized expertise. Existing AI models often process the left eye (OS) and right eye (OD) independently, failing to account for high interocular correlation and potential asymmetry. Furthermore, clinical markers like Spherical Equivalent (SE) and Axial Length (AL) are often predicted in isolation despite their biological relationship.

UWF fundus images of both eyes with clinical measurements (SE and AL).

Methods

The OUCopula framework addresses these challenges through two main architectural innovations integrated into a standard CNN backbone.

Bi-Channel Architecture with Residual Adapters

Instead of separate models for each eye, OUCopula utilizes a shared backbone network to extract common features. To account for interocular asymmetry, the authors introduced "Residual Adapters"—small, eye-specific modules that capture the unique characteristics of each eye without the computational overhead of dual full-scale models.

Copula-Likelihood Loss

The core innovation is the use of a Gaussian copula to model the joint distribution of prediction residuals for SE and AL across both eyes. Unlike standard Mean Squared Error (MSE) loss, which assumes independence between errors, the copula-likelihood loss forces the model to learn the underlying correlation structure between the four output labels (OS-SE, OS-AL, OD-SE, OD-AL).

Overview of the OUCopula framework (a) and the detailed bi-channel architecture with residual adapters (b).

Dataset and Training

The study employed a dataset of 5,228 UWF fundus images from 2,614 patients, captured using the Optomap Daytona system. Each image was graded by experienced ophthalmologists and included continuous measurements for SE and AL. The training process followed a three-stage approach: a warm-up phase for the backbone and adapters, estimation of correlation parameters, and final joint training using the copula-enhanced loss.

Results

OUCopula consistently outperformed baseline models across various metrics:

  • Overall Accuracy: The framework achieved a 7.18% improvement in total MSE compared to a standard ResNet backbone.
  • Inter-eye Synergy: Models utilizing both eyes significantly outperformed single-eye counterparts, validating the importance of interocular correlation.
  • Balanced Prediction: The copula loss ensured balanced improvements across both SE (~8.42%) and AL (~6.69%) for the left eye, respecting the biological constraints between these variables.

Conclusion

The OUCopula framework demonstrates that joint modeling of both eyes and their clinical markers significantly enhances the accuracy of myopia screening. By combining shared feature extraction with specialized adapters and a statistically grounded loss function, the model provides a robust tool for automated clinical assessment in UWF imaging.

Summary updated

Abstract

Myopia screening using cutting-edge ultra-widefield (UWF) fundus imaging is potentially significant for ophthalmic outcomes. Current multidisciplinary research between ophthalmology and deep learning (DL) concentrates primarily on disease classification and diagnosis using single-eye images, largely ignoring joint modeling and prediction for Oculus Uterque (OU, both eyes). Inspired by the complex relationships between OU and the high correlation between the (continuous) outcome labels (Spherical Equivalent and Axial Length), we propose a framework of copula-enhanced adapter convolutional neural network (CNN) learning with OU UWF fundus images (OUCopula) for joint prediction of multiple clinical scores. We design a novel bi-channel multi-label CNN that can (1) take bi-channel image inputs subject to both high correlation and heterogeneity (by sharing the same backbone network and employing adapters to parameterize the channel-wise discrepancy), and (2) incorporate correlation information between continuous output labels (using a copula). Solid experiments show that OUCopula achieves satisfactory performance in myopia score prediction compared to backbone models. Moreover, OUCopula can far exceed the performance of models constructed for single-eye inputs. Importantly, our study also hints at the potential extension of the bi-channel model to a multi-channel paradigm and the generalizability of OUCopula across various backbone CNNs.
BibTeX
@inproceedings{li2024oucopula,
  author = {Yang Li and Qiuyi Huang and Chong Zhong and Danjuan Yang and Meiyan Li and Alan H. Welsh and Aiyi Liu and Bo Fu and Catherine C. Liu and Xingtao Zhou},
  title = {{OUCopula: Bi-Channel Multi-Label Copula-Enhanced Adapter-Based CNN for Myopia Screening Based on OU-UWF Images}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure