Retinal layer segmentation of macular OCT images using boundary classification

Abstract

Optical coherence tomography (OCT) has proven to be an essential imaging modality for ophthalmology and is proving to be very important in neurology. OCT enables high resolution imaging of the retina, both at the optic nerve head and the macula. Macular retinal layer thicknesses provide useful diagnostic information and have been shown to correlate well with measures of disease severity in several diseases. Since manual segmentation of these layers is time consuming and prone to bias, automatic segmentation methods are critical for full utilization of this technology. In this work, we build a random forest classifier to segment eight retinal layers in macular cube images acquired by OCT. The random forest classifier learns the boundary pixels between layers, producing an accurate probability map for each boundary, which is then processed to finalize the boundaries. Using this algorithm, we can accurately segment the entire retina contained in the macular cube to an accuracy of at least 4.3 microns for any of the nine boundaries. Experiments were carried out on both healthy and multiple sclerosis subjects, with no difference in the accuracy of our algorithm found between the groups.

Keywords: (100.0100) Image processing; (170.4470) Ophthalmology; (170.4500) Optical coherence tomography.

Fig. 1

(a) A typical retinal OCT image (B-scan) enlarged with the layers labeled on the right-hand side. Every B-scan consists of a set of vertical scan lines (A-scans). The fovea is characterized by a depression in the surface of the retina where the top five (inner) layers are absent. (b) A fundus image with lines overlaid representing the locations of every B-scan within a volume. The red line corresponds to the B-scan in (a).

Fig. 2

Flowchart of our algorithm. The RF+CAN result refers to the segmentation using the random forest (RF) boundary classification output with a Canny-inspired boundary tracking algorithm, while RF+GS refers to the result of using the RF output with an optimal graph search (GS) algorithm.

Fig. 3

Row-wise: Shows two B-scans from within the same volume (a) with the original intensities, (b) after intensity normalization, (c) with the detected retinal boundary, and (d) after flattening.

Fig. 4

Example images of the different types of features used by the classifier: (a) the relative distance between the bottom and top boundary with contour lines overlaid, (b) the average gradient in a neighborhood below each pixel, and anisotropic Gaussian (c) first and (d) second derivatives oriented at −10 (top), 0 (center), and 10 (bottom) degrees from the horizontal.

Fig. 5

An example of the probabilities for each boundary generated as the output of the random forest classifier. The probabilities are shown for each boundary, starting from the top of the retina to the bottom, going across each row.

Fig. 6

A plot of the mean absolute error across all boundary points vs. the number of subjects, N_s, used in training the classifier. For each value of N_s, the experiment was repeated with a random set of subjects ten times. Averages are across these ten trials and error bars represent one standard deviation.

Fig. 7

(a,b) Images of the mean absolute error (μm) of each boundary at each pixel for the RF+CAN and RF+GS algorithms, respectively, with (c,d) the corresponding standard deviation of the errors. Averages are taken over all subjects and all cross-validation runs (280 values).

Fig. 8

Box and whisker plots of the mean absolute errors for every subject used in this study. Subjects are ordered by diagnosis and then age (increasing from left to right within each diagnostic group). A total of 49 data points were used to generate each subject’s plot, with each data point representing the error of a particular B-scan averaged across all cross-validation runs. For each subject, the red line represents the median absolute error and the edges of the box correspond to the 25th and 75th percentile of the error. All points lying outside of the whiskers are greater than 1.5 times the interquartile range.

Fig. 9

Two B-scan images from two different subjects are shown with the resulting boundaries from each of the 10 cross-validation runs overlaid. Each boundary is represented by a different color with the manual delineation shown atop the other boundaries in black. Therefore, if the color is not visible at a particular point, the automatic and manual segmentation are in agreement.

Fig. 10

The template for the sectors of the macula overlaid on a fundus image. The dashed square surrounding the template represents the imaged area. The concentric circles are centered on the geometric center of the OCT volume and have diameters of 1 mm, 3 mm, and 6 mm.