Detection of corneal arcus using rubber sheet and machine learning methods

Corneal Arcus (CA) is a sediment accumulation that occurs by the production of lipid (i.e. cholesterol) in the ocular eye. It is associated with hyperlipidemia caused by abnormal lipids present in the blood vessels. It occurs around the cornea with 0.3 - 1 mm wide in the iris-sclera region. The appearance of it looks like a yellowish-white ring around the cornea. This condition often occurs among older people, but in the case with young people, it is risky which associated to heart problems or stroke. In a health examination, usually an ophthalmologist or doctor who finds the patient has a CA sign, will ask them to do further treatment such as blood test. This is to ensure cholesterol (lipid) in their blood is normal or not. This procedure requires a small amount of blood taken from the blood vessels in the patient's arms. It is slightly painful, requires cost and time, besides the patient should fast for 12 hours before the test can be done. The current work for CA’s classification only focusing in the entire area of iris segmentation. This research is focusing on better ROI for iris segmentation by reducing the unwanted area, in order to maximize the useful region contain the CA presence. The segmentation iris is transformed to rectangular shape using the Rubber Sheet method. In this research, two categories of eye’s images which are the normal, and the abnormal (i.e. CA) are used. The normal eye, dataset are taken from the eye database (i.e. UBIRIS, CASIA, and IITD). Meanwhile, the CA's eye images were acquired from the medical website and the reports (e.g. journals). For the abnormal eye, the images has been examined and confirmed by a doctor who checks the images for verified that the images are the cases of CA. The framework consists of three stages of implementation such as pre-processing, features extraction and classification. First stage (pre-processing stage) consists of segmentation and normalization of the region of interest (ROI). The second stage (i.e. feature extraction stage) extracts the features based on ROI using the grey-level co- occurrence matrix (GLCM). The last stage is the classification, where it is used to identify the presence or absence of the CA. To ensure the obtained classification results are robust and stable the cross validation (CV) technique is used. The random dataset are selected by CV in the classification process (i.e. training, testing and validation). The benchmark of the classification algorithm for CA is needed to analyze the optimal output of the algorithm. The classification algorithms such as the Lavenberg-Marquardt (LM), Bayesian regularization (BR), scaled conjugate gradient (SCG) and one model of bag-of-features (BoF) are used in this research. The elements extracted from the confusion matrix parameters (i.e. accuracy, specificity, sensitivity, AUC, precision and f-score) are used in benchmarking the optimal performance of classification algorithms. Among the three neural network classifier used, BR is the best classifier. The accuracy output can be tune up to 97.2%, sensitivity 96.56%, and specificity 97.45%. On the other hand, the BoF model produced better precision of 98.04%, sensitivity 96.23%, and specificity of 100%. Based on this result, the neural network's platform for CA classification is successfully developed using the proposed framework. The result had been improved with the classification of the CA images from another benchmarking. The system can classify between CA and normal eye with good significant results.

Text FK 2019 140 - IR.pdf Download (1MB)

Actions (login required)

View Item