Diagnóstico de esclerosis múltiple mediante análisis de registros de tomografía de coherencia óptica y redes neuronales convolucionales entrenadas con imágenes sintéticas

Resumo

Background: Multiple sclerosis (MS) is a highly disabling central nervous system disease that frequently occurs in young adults. The cause of the disease is unknown, there is no biomarker for its diagnosis and there is no cure. The McDonald criteria are used for diagnosis, based mainly on MRI evidence, cerebrospinal fluid studies and the clinical status of the patient. However, it is convenient to investigate new biomarkers that allow a reliable and non-invasive diagnosis in the early stages of the disease, thus allowing the use of disease-modifying treatments, since this means a better evolution of the patients. Objectives: The general objective of this doctoral thesis is to investigate new methods of processing and classifying thickness images of different retinal structures obtained by sweptsource optical coherence tomography (SS-OCT) to achieve an early diagnosis of MS. Methods: Thickness images of the following retinal structures: whole retina, RNFL, GCL+, GCL++ and choroid, acquired by SS-OCT equipment, are available in a database consisting of 48 control subjects and 48 newly diagnosed MS patients. The Relieff method of feature categorization is used to identify the structures and regions with the highest discriminant capacity. A Convolutional Neural Network (CNN) is used as a classifier, and to avoid overfitting problems, synthetic images are generated with Generative Adversarial Networks. The validation of the classification methods is performed by cross-validation leaving one out. Results: There is no significant difference between the control group and the group of patients neither in age nor in gender distribution. The patients have had a recent diagnosis (7.35 ± 1.95 months). The application of the Relieff method detects that the three structures with the highest discriminant capacity are GCL+, GCL++ and the thickness of the complete retina. Using Generative Adversarial Networks, 100 synthetic SS-OCT images of control subjects and 100 SSOCT images of MS patients are generated. Using the original images in the CNN classifier an accuracy of 0.968 is obtained; in filtered images with the Relieff method the accuracy is 1.0 and using the synthetic images for RNC training is also 1.0. If only 50% of the original images are available, the advantage of having synthetic data for the CNN training is proven: the accuracy increases from 0.66 to 0.96. Conclusions: Neuroretinal structural alterations in the early stages of MS are suitable for implementing a diagnostic aid system using a convolutional neural network with an excellent level of accuracy.