Fluid-SegNet: Multi-dimensional loss-driven Y-Net with dilated convolutions for OCT B-scan fluid segmentation

Abstract

Optical Coherence Tomography (OCT) is a widely utilized imaging modality in clinical ophthalmology, particularly for retinal imaging. B-scan is a two-dimensional slice of the OCT volume. It enables high-resolution cross-sectional visualization of retinal layers, facilitating the analysis of retinal structure and the detection of pathological features such as fluid regions. Accurate segmentation of these fluid regions is crucial not only for determining appropriate treatment dosages but also serves as a foundation for the development of automated retinal disease diagnosis systems and visual acuity prediction models. However, the segmentation of fluid regions from OCT B-scans poses two major challenges: (1) the difficulty in delineating fine details and small fluid regions, and (2) the heterogeneity of fluid regions, which often leads to under-segmentation. This study introduces Fluid-SegNet, a novel deep learning-based segmentation framework designed to enhance the accuracy of fluid region segmentation in OCT B-scans. The proposed method is evaluated on three public datasets, UMN, AROI, and OIMHS. achieving mean Dice of 0.8725, 0.6967, and 0.8020, respectively. These results highlight the effectiveness and robustness of Fluid-SegNet in segmenting fluid regions across varied datasets and imaging conditions. Compared to existing methods, Fluid-SegNet effectively addresses the two aforementioned challenges. The source code for Fluid-SegNet is publicly available at: https://github.com/xuexiaozhong/Fluid-SegNet.