Assessment of the ISNT rule on publicly available datasets

Abstract

The ISNT rule is a technique that has been used to detect glaucoma from fundus im-ages. The rule states that for a healthy fundus image, the segmented optic disc can be divided into four neuro-retina rim quadrants namely; the Inferior, Superior, Nasal and Temporal neuro- retina rims. The Inferior is the widest followed by the Superior then the Nasal. The Temporal quadrant is the least. However, since the advent of the rule there have been several experiments that prove the inefficiency of the rule to diagnose glaucoma while other experiments argue that the rule is efficient. Experiments carried out by individuals were done using dataset sourced by the individuals not on publicly available fundus datasets. This makes the experiments not easily reproducible. This work assesses the ISNT rule using the RIM-ONE v3 dataset and the DRISHTI-GS dataset which are both publicly available datasets. The performance of the ISNT rule on the datasets is compared with the performance of a trained Extreme Gradient Boost classifier (XGB). The results show that the XGB classifier outperforms the ISNT rule and its variant. The ISNT rule demonstrated a random performance on the databases used.