Automated end-to-end Architecture for Retinal Layers and Fluids Segmentation on OCT B-scans

Age-related macular degeneration (AMD) is a degenerative retina condition that causes notable visual impairment in the central area of the visual field during its advanced stages. Manual segmentation of the retina layers and fluids in Optical coherence tomography (OCT) images is a dominant clinical practice but is time-consuming and labor-intensive due to the complexity of the images. Automated segmentation methods are necessary to facilitate efficient and quantitative assessment of the retina. However, the imbalanced distribution of classes or structures within OCT images, especially for smaller targets like intra-retinal fluids or specific layers, poses a challenge for automated segmentation methods. Algorithms may have difficulty accurately identifying and segmenting these smaller structures. The primary objective of this research is to develop segmentation models that can adapt to different pathologies, and stages of AMD disease progression while maintaining generalizability. The aim is to overcome challenges related to class imbalance, noise artifacts, intensity variations, and the scarcity of annotated OCT scans while ensuring precise and efficient segmentation. To achieve our aim, we propose a new architecture inspired by U-Net + + architecture and a Spatially adaptive denormalization Unit with a class-guided module. The segmentation is performed in two stages, to segment the layers and fluids. The proposed architecture results demonstrate significant improvement in AMD segmentation. It effectively handles class imbalance, leading to more accurate and reliable segmentation results, especially in the imbalanced classes with a Dice score of Pigment Epithelial Detachment (PED), Subretinal Fluid (SRF), Intraretinal Fluid (IRF) were 0.897, 0.927, and 0.787, respectively. Comparative analysis with other cutting-edge architectures highlights the strengths of our approach in addressing the challenges of automatic segmentation in medical images.