Detecting abnormality in optic nerve head images using a feature extraction analysis

Abstract

Imaging and evaluation of the optic nerve head (ONH) plays an essential part in the detection and clinical management of glaucoma. The morphological characteristics of ONHs vary greatly from person to person and this variability means it is difficult to quantify them in a standardized way. We developed and evaluated a feature extraction approach using shift-invariant wavelet packet and kernel principal component analysis to quantify the shape features in ONH images acquired by scanning laser ophthalmoscopy (Heidelberg Retina Tomograph [HRT]). The methods were developed and tested on 1996 eyes from three different clinical centers. A shape abnormality score (SAS) was developed from extracted features using a Gaussian process to identify glaucomatous abnormality. SAS can be used as a diagnostic index to quantify the overall likelihood of ONH abnormality. Maps showing areas of likely abnormality within the ONH were also derived. Diagnostic performance of the technique, as estimated by ROC analysis, was significantly better than the classification tools currently used in the HRT software - the technique offers the additional advantage of working with all images and is fully automated.

Keywords: (100.2960) Image analysis; (100.4993) Pattern recognition, Baysian processors; (100.7410) Wavelets; (150.1835) Defect understanding; (170.4470) Ophthalmology; (170.4580) Optical diagnostics for medicine; (170.5755) Retina scanning.

Fig. 1

An example of HRT topographic (a) and reflectance (b) images. The ONH is conventionally divided into six sectors indicated on the reflectance image.

Fig. 2

Illustration of adaptive shift-invariant wavelet packet. The value of Shannon entropy is indicated on the branches of each wavelet decomposition with the format of ‘lower level entropy: summation of higher level entropy’. The crosses indicate that the decomposition is prevented due to the increased information cost.

Fig. 3

Illustrating the number of images that select each detail coefficient in the topography and reflectance images. Brighter pixels indicate more important coefficients selected by a larger number of images. The contrast is normalized in each image for visualization purpose.

Fig. 4

The scatter plot of the first two most significant features of the ONH.

Fig. 5

The scatter plot of the first and second most significant features from those images with a diagnosis in the training data set. The axes from KPCA and PCA are in arbitrary units. The histogram of glaucoma (red) and healthy (blue) data in each feature are shown on the top and right of the scatter plot. Note that the diagnosis information was unknown during feature extraction and it is only displayed here for visualization and demonstrative purpose.

Fig. 6

ROC curve for SAS, MRA and GPS analysis. The unity line corresponds to random classification.

Fig. 7

ROC curve of SAS derived from the topography, reflectance and combined features.

Fig. 8

Examples of localized abnormality maps for ten eyes with a clinical diagnosis of glaucoma. The topography, reflectance and abnormality map are displayed for each ONH. SAS scores are given in the lower left corner of each topography image. Difference between the measured topography and reference healthy topography is represented by a heat map. The transparency of the abnormality heat map corresponds to the amount of change and a scale is shown on the margin of the abnormality map.