[Paper Review] Improving Robustness using Joint Attention Network For Detecting Retinal Degeneration From Optical Coherence Tomography Images

Sharif Amit Kamran, Alireza Tavakkoli+1 more|arXiv (Cornell University)|May 16, 2020

Retinal Imaging and AnalysisMedicine28 references30 citations

TL;DR

This paper proposes a joint-attention-network architecture that enhances robustness in retinal degeneration detection from optical coherence tomography (OCT) images by combining supervised classification with unsupervised attention map learning. By integrating a dual adaptive loss that balances classification and reconstruction, the method preserves disease-specific spatial features, significantly improving accuracy and generalization on unseen datasets across multiple state-of-the-art models.

ABSTRACT

Noisy data and the similarity in the ocular appearances caused by different ophthalmic pathologies pose significant challenges for an automated expert system to accurately detect retinal diseases. In addition, the lack of knowledge transferability and the need for unreasonably large datasets limit clinical application of current machine learning systems. To increase robustness, a better understanding of how the retinal subspace deformations lead to various levels of disease severity needs to be utilized for prioritizing disease-specific model details. In this paper we propose the use of disease-specific feature representation as a novel architecture comprised of two joint networks -- one for supervised encoding of disease model and the other for producing attention maps in an unsupervised manner to retain disease specific spatial information. Our experimental results on publicly available datasets show the proposed joint-network significantly improves the accuracy and robustness of state-of-the-art retinal disease classification networks on unseen datasets.

Motivation & Objective

• To address the challenge of low robustness in deep learning models for retinal disease classification when applied to real-world OCT images.
• To improve model generalization by preserving disease-specific spatial features lost during feature encoding in standard CNNs.
• To enable transferability across diverse OCT imaging systems without retraining by leveraging unsupervised attention learning.
• To develop a plug-and-play module that enhances existing CNN architectures for clinical deployment.

Proposed method

• Introduces a joint network with three components: a supervised encoder, an unsupervised decoder, and a supervised classifier.
• Uses skip connections to perform element-wise summation of encoder and decoder features at each stage to preserve spatial information.
• Employs a dual adaptive loss function: L = φ × Ls + (1−φ) × Lu, where Ls is cross-entropy loss for classification and Lu is mean-squared error for reconstruction.
• The hyper-parameter φ is tuned during training to balance the importance of classification and reconstruction losses.
• The unsupervised decoder generates attention maps without requiring ground truth labels, enabling domain-independent spatial feature learning.
• The architecture is integrated as a plug-in module with existing CNNs such as ResNet50, MobileNet-v2, and OpticNet-71.

Experimental results

Research questions

• RQ1 Can joint supervised and unsupervised learning improve the robustness of retinal disease classification models on unseen OCT datasets?
• RQ2 To what extent does preserving spatial feature representations through an unsupervised autoencoder component enhance model generalization?
• RQ3 How does the dual adaptive loss mechanism affect the trade-off between classification accuracy and spatial feature reconstruction?
• RQ4 Can the proposed joint-attention module be effectively transferred across different OCT imaging systems and disease types?

Key findings

• On the Srinivasan2014 dataset, the joint-attention-network achieved 100% accuracy, specificity, and sensitivity when attached to ResNet50-v1, outperforming the original model by 5.08% in all metrics.
• When tested on the unseen OCT2017 dataset, the joint-attention-network improved accuracy from 83.40% (ResNet50-v1) to 92.40%, a 9.0% relative gain.
• For MobileNet-v2, accuracy increased from 93.80% to 95.60% on OCT2017, with a 1.8% improvement in specificity and sensitivity.
• The method demonstrated consistent performance gains across all tested architectures, indicating strong transferability and robustness.
• The unsupervised decoder achieved effective reconstruction without labels, validating the effectiveness of attention map learning in preserving spatial information.
• The dual adaptive loss mechanism enabled dynamic gradient updates, with optimal φ values found to improve convergence and performance.