Automatic Segmentation of Photoreceptors in AOSLO Images Using Graph Theory and Dynamic Programming

Abstract

 

Purpose

 

The adaptive optics scanning light ophthalmoscope (AOSLO) has been a key instrument for analyzing the photoreceptor mosaic and revealing subclinical ocular pathologies. However, manual identification of photoreceptors is subjective and labor intensive. In this work, we have developed an algorithm to automatically segment and identify cone photoreceptors in AOSLO images and validated its performance against a state-of-the-art algorithm.

 

Methods

 

We extended our segmentation framework based on graph theory and dynamic programming (GTDP) to segment cone photoreceptors [1]. We used local maxima operations to obtain pilot cone location estimates and transformed each cone into the quasi-polar domain to segment and more precisely locate each cone. To validate our algorithm, we compared our GTDP algorithm to: 1) the fully automatic Garrioch implementation of the Li & Roorda method [2], and 2) the semi-automatic method from [2], where any missed cones from the Garrioch method were added manually to create the gold standard. We utilized the same dataset as in [2], which consisted of 10 repeated images in 4 parafoveal locations captured across 21 patients (840 images total). Ref: (1) Chiu et al, BOE, Vol 3, 2012. (2) Garrioch et al, Optom Vis Sci, Vol 89, 2012.

 

Results

 

Individual cones located by our GTDP method were matched with the gold standard and compared to the Garrioch method, as shown in Table 1 and Figure 1, indicating that the proposed GTDP method improved the cone detection rate of the Garrioch method (1.7 vs. 5.5% miss rate).

 

Conclusions

 

The GTDP method proposed here was able to achieve a higher detection rate compared to the state-of-the-art technique. Overall, these results are highly encouraging for reducing the time and manpower required to identify cones in ophthalmic studies.

 

 

Table 1. Cone identification performance of fully automatic methods compared to the gold standard across all 840 images. Correctly identified cones were detected in both the fully automatic and gold standard techniques, missed cones were only present in the gold standard, and additional cones were not present in the gold standard result.

 

 

Figure 1. Mean performance of the fully automatic cone identification algorithms. a) Original AOSLO image of the cone mosaic in log scale. b) Garrioch result (yellow: correctly identified; green: missed). c) GTDP result (magenta: correctly identified; green: missed; blue: added).

 

Keywords: 549 image processing • 648 photoreceptors