A COMPARISON OF DIFFERENT CLASSIFICATION STRATEGIES IN MEDICAL IMAGES OF SPECULAR MICROSCOPY TO DETECT GUTTAE IN EARLY STAGES OF FUCHS’ DYSTROPHY

Resumo

Fuchs’ endothelial dystrophy, or Fuchs‘ dystrophy, is a slowly progressive corneal disease
that usually affects both eyes. Although in many cases early signs of the disease can be seen in people
aged 20-30, the disease rarely affects vision until the person reaches the age 50-60. The tests to diagnose
the disease are: Biomicroscopy, and Specular Microscopy. In both cases, it is possible to find the
morphological changes characteristics of the disease. In this context, this paper compares the use of
three distinct machine learning techniques to perform the Fuchs‘ dystrophy diagnosis on Specular
Microscopy images in order to reduce the time spent in a manual analysis of the specialist. The
approaches used in this work were: Convolutional Neural Networks (CNN); Support Vector Machines
(SVM) with features extracted from Histogram of Oriented Gradients (HOG) and by Speeded Up Robust
Features (SURF). The dataset consists of 123,200 images of both eyes of different people, obtained over
9 years at Hospital Evangélico de Vila Velha, Espírito Santo, Brazil. Due to the absence of labels in the
original dataset, only 2400 images were analyzed and labeled with the help of a specialist. In this subset,
only in 1165 exams the Fuchs’ dystrophy is present. A cross-validation approach using 10-folds was
performed and the results were evaluated through the accuracy, area under the Receiver Operating
Characteristic (ROC) curve, precision, recall and F1 score metrics with the CNNs outperforming the
other methods.