Disc-aware Ensemble Network for Glaucoma Screening from Fundus Image (1805.07549v1)

Abstract: Glaucoma is a chronic eye disease that leads to irreversible vision loss. Most of the existing automatic screening methods firstly segment the main structure, and subsequently calculate the clinical measurement for detection and screening of glaucoma. However, these measurement-based methods rely heavily on the segmentation accuracy, and ignore various visual features. In this paper, we introduce a deep learning technique to gain additional image-relevant information, and screen glaucoma from the fundus image directly. Specifically, a novel Disc-aware Ensemble Network (DENet) for automatic glaucoma screening is proposed, which integrates the deep hierarchical context of the global fundus image and the local optic disc region. Four deep streams on different levels and modules are respectively considered as global image stream, segmentation-guided network, local disc region stream, and disc polar transformation stream. Finally, the output probabilities of different streams are fused as the final screening result. The experiments on two glaucoma datasets (SCES and new SINDI datasets) show our method outperforms other state-of-the-art algorithms.

• The paper introduces a novel Disc-aware Ensemble Network that integrates multi-scale deep learning streams to enhance glaucoma screening accuracy from fundus images.
• The methodology leverages a global image stream, a segmentation-guided network, a localized disc region stream, and a disc polar transformation stream to capture both overall and fine-grained features.
• Experimental evaluations on SCES and SINDI datasets yield high AUC scores (0.9183 and 0.8173), demonstrating robust performance for clinical glaucoma screening.

Overview of "Disc-aware Ensemble Network for Glaucoma Screening from Fundus Image"

The paper presents a novel approach to glaucoma screening using fundus images by introducing a Disc-aware Ensemble Network (DENet). This method overcomes the limitations of conventional measurement-based methods that predominately rely on segmentation accuracy by directly leveraging image-relevant information through deep learning techniques.

Methodology

The proposed model, DENet, features an intricate architecture that integrates different streams for efficient glaucoma screening:

1. Global Image Stream: This stream uses a Residual Network (ResNet-50) to process the entire fundus image, capturing global structural information. The deep architecture facilitates learning complex image representations, crucial for effective classification.
2. Segmentation-guided Network: Based on a U-shape convolutional network, this stream localizes the optic disc region within the fundus image. Its dual role involves both disc area segmentation and subsequent utilisation of segmentation-inspired representations for glaucoma screening.
3. Local Disc Region Stream: This stream operates on the optic disc cropped from the global image. It benefits from detailed local features, enhancing glaucoma detection accuracy by focusing on fine-grained visual cues inherent in the disc region.
4. Disc Polar Transformation Stream: Extending the disc region’s examination, this stream transforms it into polar coordinates. This transformation highlights structural deformities through geometric enlargement, facilitating detailed scrutiny of disc and cup morphologies.

The combination of these streams results in a network that benefits from both local and global contextual information. The ensemble employs straightforward averaging for obtaining the final glaucoma screening result, balancing computational efficiency and performance.

Experimental Findings

The effectiveness of DENet is validated through tests on SCES and SINDI datasets. The results indicate superior performance over state-of-the-art algorithms, demonstrating robust screening capabilities (evidenced by AUC scores of 0.9183 for SCES and 0.8173 for SINDI). These findings underscore DENet’s capability to circumvent traditional pitfalls associated with mere segmentation accuracy dependency and highlights its potential applicability in practical screening scenarios.

Implications and Future Directions

The reduction in reliance on segmentation precision allows for notable advancement in automated glaucoma screening, particularly apt for handling images with low contrast or pathological uncertainties. This research implies substantial improvements in screening efficiencies in clinical settings by integrating DENet into diagnostic systems, potentially leading to earlier intervention and improved patient outcomes.

For further development, there is scope to refine the ensemble strategies, possibly incorporating more sophisticated decision fusion mechanisms or dynamically adjusting stream emphasis based on individual image characteristics. Future work may also explore adapting this model for other ocular pathologies, broadening its application in comprehensive ophthalmic diagnoses.

In summary, this paper contributes significantly to the domain of medical image analysis and computer-assisted diagnosis by presenting a methodologically hybrid yet computationally feasible approach for glaucoma screening.