Generalist Segmentation Algorithm for Photoreceptors Analysis in Adaptive Optics Imaging

TL;DR

This work tackles automated cone segmentation in confocal AOSLO retinal images, addressing the labor-intensive labeling bottleneck with a semi-automatic, human-in-the-loop DL framework. The approach combines Voronoi-based masking, an attention-augmented U-Net that predicts gradient fields and region membership, and center-of-mass extraction to locate cone centers, trained on 20 batches from 18 participants across $0^{\circ}$–$2^{\circ}$ eccentricities. It achieves high F1 scores of $0.968$, $0.958$, and $0.954$ for $0^{\circ}$, $1^{\circ}$, and $2^{\circ}$, respectively, outperforming StarDist and Cellpose while reducing labeling effort by requiring only a fraction of labeled cones and iteratively refining annotations. The method holds practical significance for ophthalmology by enabling scalable, accurate cone density analyses and potential rod detection, with code made available for reproducibility.

Abstract

Analyzing the cone photoreceptor pattern in images obtained from the living human retina using quantitative methods can be crucial for the early detection and management of various eye conditions. Confocal adaptive optics scanning light ophthalmoscope (AOSLO) imaging enables visualization of the cones from reflections of waveguiding cone photoreceptors. While there have been significant improvements in automated algorithms for segmenting cones in confocal AOSLO images, the process of labelling data remains labor-intensive and manual. This paper introduces a method based on deep learning (DL) for detecting and segmenting cones in AOSLO images. The models were trained on a semi-automatically labelled dataset of 20 AOSLO batches of images of 18 participants for 0$^{\circ}$, 1$^{\circ}$, and 2$^{\circ}$ from the foveal center. F1 scores were 0.968, 0.958, and 0.954 for 0$^{\circ}$, 1$^{\circ}$, and 2$^{\circ}$, respectively, which is better than previously reported DL approaches. Our method minimizes the need for labelled data by only necessitating a fraction of labelled cones, which is especially beneficial in the field of ophthalmology, where labelled data can often be limited.

Figures (8)

• Figure 1: Two examples of confocal AOSLO images with labeled cone centers using the existing semi-automatic segmentation method li2007automated followed by refinement by a medical expert.
• Figure 2: Pipeline of the method. Voronoi algorithm is applied to initially semi-automatically annotated cones to obtain the masks. Then, a segmentation model was trained, which generates segmentation masks for unlabeled AOSLO images. The center of mass function is applied to get the centers of the cells from segmentation masks. After each iteration step, the model was evaluated using the test subset. A manual correction step is involved in the pipeline to improve the annotations of the segmentation model of initially unlabeled images.
• Figure 3: Application of Voronoi algorithm on the labeled AOSLO images: (a) example of the original image; (b) segmented image.
• Figure 4: Model overview. (a) Transformation from the center of the cell to a gradient vector field using the Voronoi algorithm. (b) U-Net model with additional Attention-augmented module.
• Figure 5: Evaluation of the proposed method on first (a) and second (b) iterations on a 0$^{\circ}$ test sample. Green circles correspond to the ground truth cone centers, and red to the predicted centers. Yellow squares show the False Negative predictions of the model.