Attention Based Glaucoma Detection: A Large-scale Database and CNN Model (1903.10831v3)

Abstract: Recently, the attention mechanism has been successfully applied in convolutional neural networks (CNNs), significantly boosting the performance of many computer vision tasks. Unfortunately, few medical image recognition approaches incorporate the attention mechanism in the CNNs. In particular, there exists high redundancy in fundus images for glaucoma detection, such that the attention mechanism has potential in improving the performance of CNN-based glaucoma detection. This paper proposes an attention-based CNN for glaucoma detection (AG-CNN). Specifically, we first establish a large-scale attention based glaucoma (LAG) database, which includes 5,824 fundus images labeled with either positive glaucoma (2,392) or negative glaucoma (3,432). The attention maps of the ophthalmologists are also collected in LAG database through a simulated eye-tracking experiment. Then, a new structure of AG-CNN is designed, including an attention prediction subnet, a pathological area localization subnet and a glaucoma classification subnet. Different from other attention-based CNN methods, the features are also visualized as the localized pathological area, which can advance the performance of glaucoma detection. Finally, the experiment results show that the proposed AG-CNN approach significantly advances state-of-the-art glaucoma detection.

• The paper introduces an attention-based CNN that integrates three specialized subnets for precise glaucoma detection from fundus images.
• It employs attention mechanisms derived from simulated eye-tracking to generate and refine attention maps, enhancing pathological area localization.
• Using the comprehensive LAG database, the model achieves 95.3% accuracy and robust external validation, highlighting its clinical potential.

Overview of the Attention-Based CNN for Glaucoma Detection

The paper presents a compelling approach to glaucoma detection utilizing an attention-based convolutional neural network (AG-CNN). This paper is distinguished by its integration of attention mechanisms in medical imaging, an area where such approaches have previously been underutilized. The primary contributions are the development of an AG-CNN architecture tailored for detecting glaucoma from fundus images and the creation of a comprehensive dataset (LAG database) encompassing 5,824 annotated fundus images, complete with attention maps obtained through simulated eye-tracking experiments.

The AG-CNN model is structured into three subnets: attention prediction, pathological area localization, and glaucoma classification. This architecture facilitates the model's focus on salient image regions, effectively reducing redundancy. The attention prediction subnet generates attention maps that are critical for identifying areas of clinical significance within fundus images. The pathological area localization subnet refines these maps to ensure they correspond accurately to regions indicative of glaucoma. Finally, the classification subnet applies these insights to achieve high-accuracy predictions of glaucoma presence.

Experimental Results and Discussion

The AG-CNN was trained and tested on the authors' LAG database and externally validated on the RIM-ONE database, a public dataset, to assess generalization capabilities. The performance metrics of AG-CNN are robust, demonstrating an accuracy of 95.3% on their test dataset and substantial improvements over competing models. Specificity and sensitivity metrics were closely aligned, underscoring balanced detection capabilities. The receiver operating characteristic (ROC) analyses confirm AG-CNN’s high true positive rate.

When compared to existing models, AG-CNN shows substantial improvements, particularly on sensitivity and specificity, indicating fewer instances of false negatives and maintaining a low false positive rate. This is critical in glaucoma detection, where early detection can prevent irreversible optic nerve damage.

Implications and Future Directions

The proposed AG-CNN model and the accompanying LAG database pave the way for further application of attention mechanisms within medical image analysis. Given glaucoma's complexity, ensuring the model focuses on pertinent anatomical structures—such as the optic cup, disc, and surrounding regions—is vital, and the attention mechanism addresses this necessity effectively.

Looking ahead, extensions of this research could explore the integration of additional ophthalmic datasets to reinforce and potentially enhance the model's robustness. Another avenue for advancement is the incorporation of other ophthalmic conditions to establish a multi-disease detection framework within ophthalmology, enhancing clinical applicability. Moreover, harnessing explainability techniques to further elucidate model decisions could bolster clinical confidence in these automated approaches.

In conclusion, this paper highlights a methodologically innovative approach within medical imaging for glaucoma detection. By effectively applying an attention-driven framework, the research marks a significant step towards more accurate, efficient, and clinically interpretable automatic disease detection tools in ophthalmology.