Split bundle detection in polarimetric images of the human retinal nerve fiber layer.

Objectives: One method for assessing pathological retinal nerve fiber layer (NFL) appearance is by comparing the NFL to normative values, derived from healthy subjects. These normative values will be more specific when normal physiological differences are taken into account. One common variation is a split bundle. This paper describes a method to automatically detect these split bundles.
Methods: The thickness profile along the NFL bundle is described by a non-split and a split bundle model. Based on these two fits, statistics are derived and used as features for two non-parametric classifiers (Parzen density based and k nearest neighbor). Features were selected by forward feature selection. Three hundred and nine superior and 324 inferior bundles were used to train and test this method.
Results: The prevalence of split superior bundles was 68% and the split inferior bundles' prevalence was 13%. The resulting estimated error of the Parzen density- based classifier was 12.5% for the superior bundle and 10.2% for the inferior bundle. The k nearest neighbor classifier errors were 11.7% and 9.2%.
Conclusions: The classification error of automated detection of split inferior bundles is not much smaller than its prevalence, thereby limiting the usefulness of separate cut-off values for split and non-split inferior bundles. For superior bundles, however, the classification error was low compared to the prevalence. Application of specific cut-off values, selected by the proposed classification system, may therefore increase the specificity and sensitivity of pathological NFL detection.