PY-Net: Rethinking segmentation frameworks with dense pyramidal operations for optic disc and cup segmentation from retinal fundus images

• Published in: Biomedical signal processing and control 2023-08, Vol.85, p.104895, Article 104895
• Main Authors: Bhattacharya, Rajarshi, Hussain, Rukhshanda, Chatterjee, Agniv, Paul, Dwipayan, Chatterjee, Saptarshi, Dey, Debangshu
• Format: Article
• Language: English
• Subjects: Deep learning; Glaucoma detection; Image processing; Optic disc and cup segmentation
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2023.104895

Glaucoma is one of the leading causes of irreversible blindness across the globe. Early diagnosis is therefore essential for the provision of timely treatments and reducing the loss of vision. From the retinal fundus images, the calculation of the cup to disc ratio (CDR) is an efficient indicator to distinguish between glaucomatous and non-glaucomatous cases. Therefore, precise segmentation of optic disc and cup from the retinal images are important steps for improved diagnosis of the disease. This paper presents a robust segmentation pipeline for optic disc and cup segmentation utilizing the U-Net architecture. In the upsampling half of the model, a spatial pyramid based decoder cascaded with an intermediate decoder is introduced. To incorporate channel-wise attention throughout the framework squeeze and excite blocks are incorporated. To enhance the relevant spatial features in the feature representations of the fundus images, spatial attention module is utilized. Furthermore to achieve multi-scale context extraction modified receptive field blocks (MRFB) are added to the encoding layers of the network. In addition, the MRFBs are also introduced in the auxiliary decoder for fine-tuning of the feature representation. The performance of the proposed model has been evaluated on three publicly available retinal fundus databases, DRISHTI-GS, RIM-ONE, and REFUGE and the model has achieved a dice score of 97.10, 96.12, 96.48 respectively on the optic disc and 93.38, 92.92, 93.85 on optic cup segmentation, outperforming the state-of-the-art methods. [Display omitted] •A pipeline with auxiliary and densely-connected pyramidal decoder is proposed.•Receptive Field Blocks extract contextual features with finer details.•Spatial Attention blocks generate attention maps from low-level features.•Squeeze-and-Excite blocks scale up or scale down the features of each channel.•A linear combination of various losses is used to train the model.