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.

Motivation

As shown in Figure 1, glaucoma can be correctly detected by a CNN method when the visualized heat maps are consistent with the attention maps of ophthalmologists in glaucoma diagnosis. Otherwise, glaucoma is mislabeled by the CNN model. Therefore, it is reasonable to combine the attention mechanism in the CNN model for using fundus images to detect ophthalmic diseases.

Figure 1. Examples of glaucoma fundus images, attention maps by ophthalmologists in glaucoma diagnosis and visualization results of a CNN model by an occlusion experiment.

Method

The framework of AG-CNN is shown in Figure 4. As shown in Figure 4, the input to AG-CNN is the RGB channels of a fundus image, while the output is (1) the located pathological area and (2) the binary glaucoma label. Our AG-CNN has two 2 stages as follows.

• In the first stage, the ROI of glaucoma detection is learned from the attention prediction subnet, aiming to predict human attention on diagnosing glaucoma. In addition, the multiscale building block is also included in our AG-CNN model, because \textit{Finding 3} shows that ROIs for glaucoma diagnosis are of different scales.
• In the second stage, the predicted attention map is embedded in the pathological area localization subnet, and then the feature map of this subnet is visualized to locate the pathological area. Finally, the located pathological area combined with the predicted attention map is further used to mask the input and features of the glaucoma classification subnet, for outputting the binary labels of glaucoma.

Publication

1. Liu Li, Mai Xu, Hanruo Liu, et al., "A Large-scale Database and a CNN Model: Attention-based Glaucoma Detection" in IEEE Transactions on Medical Imaging, 2019.
2. Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu, "Attention Based Glaucoma Detection: A Large-scale Database and CNN Model" in Conference on Computer Vision and Pattern Recognition, 2019. [Paper][Database][poster]