Attention Based Glaucoma Detection: A Large-scale Database and CNN Model

Liu Li; Mai Xu; Xiaofei Wang; Lai Jiang; Hanruo Liu

Source-linked AI summary

Attention Based Glaucoma Detection: A Large-scale Database and CNN Model

Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, Hanruo Liu

arXiv:1903.10831v3Mar 26, 2019cs.CV

TL;DR

Glaucoma fundus images contain redundant regions, while medical image recognition has rarely incorporated human attention. The paper builds the LAG database and proposes AG-CNN to predict attention, localize pathological areas, and classify glaucoma; experiments report improved detection performance over comparison methods.

Problem
Medical image-recognition methods rarely incorporate human attention, despite substantial redundancy in fundus images for glaucoma detection.
Method
The paper establishes the LAG database and designs AG-CNN with attention prediction, pathological-area localization, and glaucoma-classification subnets.
Results
AG-CNN achieves 95.3% accuracy, 95.4% sensitivity, 95.2% specificity, and a 0.951 F2-score, outperforming two comparison methods across reported metrics.
Takeaways & Limitations
Predicted attention maps and localized pathological areas improve glaucoma-detection performance in AG-CNN.

Abstract

from arXiv · show

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.

1. Introduction

Attention mechanisms can reduce redundant visual information, making them relevant to glaucoma detection in fundus images. The paper addresses the lack of human-attention data and proposes the LAG database and AG-CNN, which also localizes pathological areas.

Attention mechanisms locate salient deep-network features to remove redundancy in computer-vision tasks.
Fundus images contain substantial redundant regions, while glaucoma-related pathology is concentrated around the optic cup, disc, blood vessels, and optic nerve.
13?
Human attention can miss pathological areas, motivating guided back propagation to refine predicted attention maps and highlight tiny critical regions.
The paper contributes the LAG database and an AG-CNN architecture that predicts attention, localizes pathological areas, and classifies binary glaucoma.

2. Medical Background

Prior glaucoma-detection methods use heuristic image-processing features or deep learning, including optic-cup-and-disc segmentation and end-to-end CNN classification. These approaches motivate an attention-based model that can address redundant fundus-image information.

Glaucoma-detection methods are broadly divided into heuristic and deep-learning approaches.
Heuristic methods extract texture, higher-order spectra, or wavelet-energy features before classification with SVM or naive Bayesian models.
Some deep-learning methods segment the optic cup and disc but assume these structures alone are related to glaucoma and lack end-to-end training.
Other work uses end-to-end CNNs for glaucoma detection, extending deep learning beyond manually engineered image features.

3. Database

The LAG database combines glaucoma labels with ophthalmologists’ attention data collected through simulated eye tracking. Its analyses examine fixation consistency, ROI concentration, and variation in ROI scale across fundus images.

3.1. Establishment: The LAG database contains 5,824 fundus images, including 2,392 positive and 3,432 negative glaucoma samples.
3.1. Establishment: Ophthalmologists’ attention regions were captured by successively clearing blurred circular regions with mouse clicks during glaucoma diagnosis.
3.1. Establishment: Each attention map aggregates ophthalmologists’ fixation coordinates with a 2D Gaussian filter whose decay follows fixation order; the experiment used four ophthalmologists and a standard deviation of 25.
3.2. Data analysis: The database analysis finds that ophthalmologists’ diagnostic regions are consistent across observers.
3.2. Data analysis: 14.3% of each fundus image was cleared on average, ranging from 11.8% to 17.8%, indicating that diagnostic attention concentrates on small regions.
3.2. Data analysis: The diagnostic ROI varies in scale across images, including both positive and negative glaucoma cases and differences associated with pathological myopia.

4. Method

AG-CNN combines human-attention prediction, pathological-area visualization, and glaucoma classification in an end-to-end architecture. Predicted and visualized regions are used to suppress redundant fundus-image information while preserving potentially useful pathological features.

Framework: AG-CNN learns glaucoma-related regions in two stages: predicting human attention, then visualizing CNN features to localize pathological areas.The localized area is subsequently used by the classification subnet.
Attention prediction subnet: The attention prediction subnet converts a 224 × 224 × 3 fundus image into a 112×112×1 gray attention map using hierarchical and multi-scale features.It processes features from four hierarchical building blocks with feature normalization, concatenates them, and applies deconvolution.
Pathological area localization subnet: The pathological area localization subnet masks inputs and intermediate features with predicted attention maps, then uses guided BP to produce a normalized pathological-area visualization map.The visualization map is down-sampled to 112 × 112 with values normalized to 0−1.
Glaucoma classification subnet: The glaucoma classification subnet uses multi-scale convolutional channels, concatenated features, and fully connected layers to output binary glaucoma labels.Its multi-scale building block uses four convolutional channels with different kernel sizes rather than a single convolutional channel.
Glaucoma classification subnet: A visualization map masks classification features and the input image using θ = 0.5, inhibiting redundant features while highlighting pathological regions.The masking combines element-wise multiplication and addition, with θ controlling the visualization map’s impact.
Loss function: AG-CNN is trained end-to-end with attention prediction, pathological-area localization, and glaucoma classification supervision combined as Loss = α · Lossa + β · Lossf + γ · Lossc.Training initially emphasizes attention prediction, then shifts emphasis toward visualization and classification losses.

5. Experiments and Results

Experiments evaluate AG-CNN for glaucoma detection, attention prediction, and pathological-area localization on LAG and RIM-ONE. AG-CNN outperforms comparison methods and benefits from its attention, localization, and multi-scale components.

Glaucoma detection: 95.3% accuracy, 95.4% sensitivity, and 95.2% specificity were achieved by AG-CNN on LAG, compared with F2-scores of 0.894 and 0.901 for the two comparison methods.AG-CNN is reported to outperform both comparison methods across all listed metrics.
Glaucoma detection: AG-CNN’s ROC curve was closer to the upper-left corner, with higher sensitivity than both comparison methods at the same specificity.The ROC comparison was conducted on the LAG testing set, and AUC was also evaluated.
Generalization: All AG-CNN metrics on RIM-ONE exceeded 0.83, and its performance was especially better in sensitivity despite being slightly below its LAG results.The comparison method showed severe performance degradation attributed to over-fitting, while AG-CNN was reported to generalize better.
Attention prediction: 0.934 average CC between predicted and ground-truth attention maps on LAG indicates accurate attention prediction, including across differently scaled RIM-ONE images.RIM-ONE lacked ground-truth attention maps, so scale robustness was assessed through the predicted maps’ localization of the optic cup and disc.
Pathological-area localization: 0.581 average CC between localized pathological areas and ophthalmologists’ attention maps showed consistency, while AG-CNN better pinpointed tiny optic-disc-edge abnormalities than CAM.The localization also covered the optic cup, disc, and retinal nerve fiber layer defects.
Ablation experiments: 4.5%, 4.3%, 4.7%, and 4.7% improvements in accuracy, sensitivity, specificity, and F2-score resulted from jointly adding predicted attention maps and pathological-area localization.Removing attention prediction caused pathological areas to spread across the fundus, and the multi-scale block further enhanced glaucoma-detection performance.

6. Conclusion

The paper proposes AG-CNN for automatic glaucoma detection and pathological-area localization using fundus images. It also establishes the LAG database and reports that predicted attention maps improve both detection and localization relative to other methods.

6. Conclusion: AG-CNN combines attention prediction, pathological-area localization, and glaucoma classification for fundus-image analysis.The model classifies glaucoma using deep features highlighted by visualized pathological-area maps based on predicted attention maps.
6. Conclusion: The LAG database contains 5,824 labeled fundus images with ophthalmologists’ attention maps for glaucoma detection.The database supports training the proposed attention-based model.
6. Conclusion: Predicted attention maps significantly improve AG-CNN’s glaucoma detection and pathological-area localization and outperform other state-of-the-art methods.This is the paper’s reported overall experimental conclusion.