Image Projection Network: 3D to 2D Image Segmentation in OCTA Images

Abstract

Optical coherence tomography angiography (OCTA) is a novel noninvasive imaging modality that allows micron-level resolution to visualize the retinal microvasculature. The retinal vessel segmentation in OCTA images is still an open problem, and especially the thin and dense structure of the capillary plexus is an important challenge of this problem. In this work, we propose a novel image magnification network (IMN) for vessel segmentation in OCTA images. Contrary to the U-Net structure with a down-sampling encoder and up-sampling decoder, the proposed IMN adopts the design of up-sampling encoding and then down-sampling decoding. This design is to capture more image details and reduce the omission of thin-and-small structures. The experimental results on three open OCTA datasets show that the proposed IMN with an average dice score of 90.2% achieves the best performance in vessel segmentation of OCTA images. Besides, we also demonstrate the superior performance of IMN in crossfield image vessel segmentation and vessel skeleton extraction.

ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model

“…Li et.al. [21] presented an image projection network, which is a novel end-to-end architecture that can achieve 3D-to-2D image segmentation in OCTA images.…”

“…Mou et al trained a deep network [1] with 30 OCTA images, but the small dataset used suggests that this method may not be universally applicable across different pathological scenarios. Li et al [21] more recently introduced an image projection network that can achieve 3Dto-2D vessel segmentation: they evaluated their method on (in bottom row), respectively. Columns 1 to 3 respectively show the original OCTA images, manually annotated vessel networks by experts, and zoomed image patches of the same size (0.5 × 0.5 mm 2 ) in both scans.…”

“…In addition, Eladawi [8] and Li et al [21] set up their segmentation frameworks based on the OCTA data within a 6 × 6 mm 2 fovea-centered field of view (FOV). It is worth noting that the scan density for the 3 × 3 mm 2 is higher than that for the 6 × 6 mm 2 scans [22], 6 × 6 mm 2 scans has relatively wider area of scan coverage, and is more like to detect the presence of pathological features, e.g., microaneurysms and non-perfusion.…”

Foveal Avascular Zone Segmentation of Octa Images Using Deep Learning Approach with Unsupervised Vessel Segmentation

“…Recently, Menglin et al [2] proposed a U-Net network with a squeeze-and-excitation (SE) block [14] and used post processing to further improve FAZ segmentation performance. Also, Mingchao et al [15] proposed an image projection network that enables the network to generate 2D segmentation from 3D voxels. However, these methods only show higher performance on its own dataset.…”

Fast 3-dimensional estimation of the Foveal Avascular Zone from OCTA

“…A detailed review can be found in [2], [3]. Efforts to improve and automatize the measurement are ongoing, with new publications exploiting the latest technologies [8]- [13]. However, only a few of these publications are based on 3D measurements from OCTA [8], [13], and no attempt to segment the FAZ in 3D has been made yet.…”