Geração de populações artificiais para treinamento de classificadores de apoio ao diagnóstico de glaucoma.

Abstract

Glaucoma is an optical neuropathy, whose progression results in visual field impairments and blindness. Due to its irreversible damages, early and correct identification is very important to control glaucoma's progression. For glaucoma diagnosis, ophthalmologists analyze patient's visual field and eyes structural data obtained by using eyes' test equipments. In order to reduce the quantity of false-negative and false-positive results, several new techniques have been developed to increment the sensitivity and specificity of glaucoma diagnostic tests. A promising approach is the use of machine learning classifiers. Classifiers based on different concepts like Decision Trees, Artificial Neural Networks, and Bayesian approach, have been developed for medical applications. Despite the availability of sophisticated algorithms for classifiers development, successful training of classifiers is highly dependent of good training data. Good data means that examples provided for classifier training should represent the many different situations found in real world. These requirements are usually accomplished if data from a large number of patients is available. However, several factors like profile of evaluated population, duration of the data acquisition activities, existence of healthcare professionals, and equipment availability, and people's commitment to the research program, restrict the size of patient's dataset. A possible approach to overcome the lack of patient's data to perform the classifier's training task is to use artificial data that represent a real population. This artificial data would be suitable for classifiers training if it has similar statistical properties of a real population. The use of artificial population will enable the creation of datasets with required number of patients, and without spending years measuring patients. It will also be possible to simulate scenarios and strategies before a long term research program starts. In this work is presented an artificial data generator named GLOR, based on a Monte Carlo method, and suitable for the training of classifiers for glaucoma diagnosis. The generated population is characterized by eyes'functional and structural data provided by Standard Automated Perimetry (SAP) and High Definition Optical Coherence Tomography (HD-OCT) instruments. The experimental results, obtained after an Artificial Neural Network training employing a population generated by GLOR comprising of 4500 normal and 500 glaucomatous individuals and evaluated by using real population data from 44 normal and 26 glaucomatous subjects, were: 87.1% for overall accuracy, 80.8% for sensitivity, 90.9% for specificity and 0.941 for the area under ROC curve. These results show that GLOR can be used as a promising approach to accelerate the development of new methods to increment sensitivity and specificity of glaucoma diagnosis.