MSFCS-Net: A Dual-Decoder Architecture for an Efficient Segmentation of Optic Disc and Optic Cup

Fundus photography significantly enhances the detection mechanism of ophthalmic conditions that affect the vision of the eye. One of these conditions, Glaucoma, which is widely known as the silent thief of vision, is due to its subtle signs in causing vision loss. Precise localisation of the Optic Cup (OC) & Optic Disc (OD) from retinal images enhances the diagnostic accuracy of Glaucoma. Thus, to develop an automated model, this research proposes a Multi-Scale Feature-driven Condition-based Segmentation Network (MSFCS-Net), an improved Deep Convolutional Neural Network (DCNN) that elevates the segmentation efficiency by learning feature maps of several sizes using a multi-scale encoder. In this paper, the Residual Depth-wise Separable Convolution Module (RDSCM) is employed to preserve spatial information, reducing the hyperparameter size to mitigate the burden of computational complexity. Furthermore, a Triple Attention Mechanism (TAM) is integrated to elevate the pertinent feature maps, enabling the network to achieve high intra-class variance between the pixel values of different classes. Additionally, a novel condition-based segmentation mechanism is proposed to control the segmentation of the OC using the predicted maps of the OD via a dual decoder of the MSFCS-Net. This model has shown a remarkable improvement in the segmentation of OD and OC regions when validated with different publicly available datasets such as REFUGE, Rim-One, and ORIGA, for accurate diagnosis of Glaucoma. The MSFCS-Net model achieved an average accuracy of 99.67, an intersection over union of 99.34 and a Dice coefficient of 99.66 on the ORIGA, outperforming the existing models with the proposed model.