Joint Optic Disc and Cup Segmentation Based on Multi-label Deep Network and Polar Transformation (1801.00926v3)

Abstract: Glaucoma is a chronic eye disease that leads to irreversible vision loss. The cup to disc ratio (CDR) plays an important role in the screening and diagnosis of glaucoma. Thus, the accurate and automatic segmentation of optic disc (OD) and optic cup (OC) from fundus images is a fundamental task. Most existing methods segment them separately, and rely on hand-crafted visual feature from fundus images. In this paper, we propose a deep learning architecture, named M-Net, which solves the OD and OC segmentation jointly in a one-stage multi-label system. The proposed M-Net mainly consists of multi-scale input layer, U-shape convolutional network, side-output layer, and multi-label loss function. The multi-scale input layer constructs an image pyramid to achieve multiple level receptive field sizes. The U-shape convolutional network is employed as the main body network structure to learn the rich hierarchical representation, while the side-output layer acts as an early classifier that produces a companion local prediction map for different scale layers. Finally, a multi-label loss function is proposed to generate the final segmentation map. For improving the segmentation performance further, we also introduce the polar transformation, which provides the representation of the original image in the polar coordinate system. The experiments show that our M-Net system achieves state-of-the-art OD and OC segmentation result on ORIGA dataset. Simultaneously, the proposed method also obtains the satisfactory glaucoma screening performances with calculated CDR value on both ORIGA and SCES datasets.

• The paper presents M-Net, a novel deep learning framework that jointly segments the optic disc and cup using a multi-label Dice loss function.
• It leverages polar transformation to remap fundus images into a polar coordinate system, improving the delineation of overlapping regions.
• Experimental results on ORIGA and SCES datasets show state-of-the-art segmentation accuracy and strong CDR-based glaucoma screening performance with AUCs of 0.85 and 0.90.

Joint Optic Disc and Cup Segmentation Based on Multi-label Deep Network and Polar Transformation

The paper "Joint Optic Disc and Cup Segmentation Based on Multi-label Deep Network and Polar Transformation" by Huazhu Fu et al. introduces a deep learning framework, M-Net, designed to address challenges in the automatic segmentation of the optic disc (OD) and optic cup (OC) from fundus images. This segmentation is pivotal for glaucoma screening, where the cup-to-disc ratio (CDR) is a critical diagnostic measure.

Methodology

The authors propose M-Net as a joint multi-label segmentation architecture that simultaneously identifies both the OD and OC, addressing limitations of prior methods that treat the tasks separately. M-Net employs a U-shape convolutional network structure with several key components: a multi-scale input layer, side-output layers, and a multi-label loss function based on the Dice coefficient.

1. Multi-scale Input Layer: This layer constructs an image pyramid that facilitates varying levels of receptive fields, enhancing the network's ability to capture distinct features at multiple scales.
2. Side-output Layers: These layers act as early classifiers by producing local prediction maps, aiding in better back-propagation and reducing the vanishing gradient problem.
3. Multi-label Loss Function: By treating OD and OC as independent labels, the network effectively handles the overlap between these regions, utilizing a Dice-based loss to accommodate class imbalance.

Furthermore, the integration of polar transformation optimizes the segmentation process by mapping images into a polar coordinate system. This transformation leverages spatial constraints and balances the proportion of the cup region, subsequently improving segmentation performance.

Experimental Results

The proposed method was evaluated on the ORIGA and SCES datasets. Notably, M-Net achieved state-of-the-art segmentation results with overlapping errors of 0.07 for OD and 0.23 for OC on the ORIGA dataset. The calculated CDR demonstrated strong glaucoma screening performance, yielding AUCs of 0.85 and 0.90 on the ORIGA and SCES datasets, respectively. These results underscore M-Net's efficacy, particularly its robustness in large-scale screening scenarios.

Implications and Future Directions

The research presents significant advancements in automated glaucoma screening. By leveraging a joint segmentation approach, the proposed framework enhances efficiency and accuracy in diagnostics, offering potential integration into clinical workflows.

Future developments could explore the applicability of this method to other ocular pathologies. Additionally, further refinement of polar transformation and incorporation of additional data augmentation techniques could yield even higher segmentation accuracy. Expanding the dataset scope and incorporating more diverse clinical images could also improve the generalizability of the model.

Overall, this paper provides a comprehensive contribution to the field of medical image analysis, particularly in the context of ophthalmology, and sets a platform for future explorations into multi-label segmentation tasks within complex medical datasets.