266
F. Shi et al.
10.4.2 Method
10.4.2.1 Method Overview
The proposed method consists of three main parts: pre-processing, classification, and
post-processing. In the pre-processing step, the SD-OCT images are first denoised
and segmented into 10 intra-retinal layers with 11 surfaces. The retina in the original SD-OCT volume is flattened, where the 11th surface (the bottom of the retinal
pigment epithelium) is used as the reference plane. The EZ region between the 7th
and 8th surfaces is extracted, which is the volume of interest (VOI) for our analysis.
In the classification step, five categories, from a total of 57 features, are extracted
for each voxel in the VOIs. Then, principle component analysis (PCA) is adopted
for feature selection. Because the disrupted voxels (the minority) in the VOIs are far
less numerous than the non-disrupted ones (the majority), it is a typical imbalanced
classification problem. To improve the performance of the classification, we apply
the following two strategies in the classification training: (1) an Adaboost algorithm
is adopted to train some weak classifiers into an integrated strong classifier at the
algorithm level; and (2) the majority samples are randomly under-sampled at the data
level. In the classifier testing step, every voxel in the VOIs is classified as disrupted or
not disrupted. In the post-processing step, the blood vessel silhouettes are identified
and excluded by a vessel detector and the isolated points are excluded by morphological operations to avoid false detections. Finally, the volume of the disrupted EZ
is calculated.
10.4.2.2 Pre-processing
Speckle noise is the main noise in OCT images, and it affects the performance of
image processing and classification. In this paper, we propose applying the bilateral
filtering [32] method for denoising because it can remove speckle noise from images
effectively while maintaining edge-like features. We have used a fast approximation
algorithm [33] to reduce the computation time without significantly impacting the
bilateral filtering result. Each B-scan (X-Z image) of the OCT images is smoothed
separately by bilateral filtering.
The filtered SD-OCT volume is then automatically segmented into 10 intra-retinal
layers using the multi-scale 3D graph-search approach [1, 2, 14, 28, 29], which
produces 11 surfaces (see Fig. 10.14). Then, all the surfaces are smoothed using thin
plate splines. The retina in the original SD-OCT volume is flattened by adjusting the
A-scans up and down in the z-direction, where the 11th surface (the bottom of the
retinal pigment epithelium) is used as a reference plane because of its robustness.
Then, the EZ regions between the 7th and 8th surfaces are extracted as the volumes
of interest (VOIs).
F. Shi et al.
10.4.2 Method
10.4.2.1 Method Overview
The proposed method consists of three main parts: pre-processing, classification, and
post-processing. In the pre-processing step, the SD-OCT images are first denoised
and segmented into 10 intra-retinal layers with 11 surfaces. The retina in the original SD-OCT volume is flattened, where the 11th surface (the bottom of the retinal
pigment epithelium) is used as the reference plane. The EZ region between the 7th
and 8th surfaces is extracted, which is the volume of interest (VOI) for our analysis.
In the classification step, five categories, from a total of 57 features, are extracted
for each voxel in the VOIs. Then, principle component analysis (PCA) is adopted
for feature selection. Because the disrupted voxels (the minority) in the VOIs are far
less numerous than the non-disrupted ones (the majority), it is a typical imbalanced
classification problem. To improve the performance of the classification, we apply
the following two strategies in the classification training: (1) an Adaboost algorithm
is adopted to train some weak classifiers into an integrated strong classifier at the
algorithm level; and (2) the majority samples are randomly under-sampled at the data
level. In the classifier testing step, every voxel in the VOIs is classified as disrupted or
not disrupted. In the post-processing step, the blood vessel silhouettes are identified
and excluded by a vessel detector and the isolated points are excluded by morphological operations to avoid false detections. Finally, the volume of the disrupted EZ
is calculated.
10.4.2.2 Pre-processing
Speckle noise is the main noise in OCT images, and it affects the performance of
image processing and classification. In this paper, we propose applying the bilateral
filtering [32] method for denoising because it can remove speckle noise from images
effectively while maintaining edge-like features. We have used a fast approximation
algorithm [33] to reduce the computation time without significantly impacting the
bilateral filtering result. Each B-scan (X-Z image) of the OCT images is smoothed
separately by bilateral filtering.
The filtered SD-OCT volume is then automatically segmented into 10 intra-retinal
layers using the multi-scale 3D graph-search approach [1, 2, 14, 28, 29], which
produces 11 surfaces (see Fig. 10.14). Then, all the surfaces are smoothed using thin
plate splines. The retina in the original SD-OCT volume is flattened by adjusting the
A-scans up and down in the z-direction, where the 11th surface (the bottom of the
retinal pigment epithelium) is used as a reference plane because of its robustness.
Then, the EZ regions between the 7th and 8th surfaces are extracted as the volumes
of interest (VOIs).