Uma abordagem para diagnóstico automático de estrabismo baseado em vídeos do exame Cover test alternado utilizando Deep Learning

• Autor(a) principal: SANTOS, Robert Douglas de Araujo
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/3844
• topic: estrabismo;; diagnóstico automático;; covert test alternado; YOLOv5.; strabismus;; automatic diagnosis;; alternate covert test;; YOLOv5.; Oftalmologia; Ciência da Computação

A condition known as strabismus occurs when one eye’s line of sight is unable to fix on the target item that both eyes are picturing; rather, when one eye fixes its gaze on the object, the other eye’s line of sight is in a different direction. Strabismus affects a growing portion of the adult population and a large proportion of children. It can cause aesthetic issues and vision loss, which can be avoided in many cases. Taking into account the characteristics of the disturb, the number of specialized doctors available to the public, and computer advancements in recent years, this study presents a computational method for detecting and diagnosing strabismus using a neural convolutional network YOLOv5 for ocular detection and occluder in digital exam videos of Cover Test. The method was carried out on two video databases, the first of which is called Dataset CV-I and contains 13 volunteer videos obtained in the hospital the Universidade Federal do Maranhão. The second basis, known as Dataset CV-II, is made up of 57 volunteer videos that were acquired in public schools in the city of São Luís, Maranhão. The results obtained by the proposed method are purchased with studies by a specialist doctor. According to the results, the proposed method correctly detects the eyes 100% of the time and correctly occluder 97% of the time in the training and testing dataset. Using the video base of sixth nerve palsy, also known as PSN, which comprises 35 movies collected in various contexts, the assertiveness of the neural network in identifying the eyes was also evaluated. accordingly, there are 19,966 eyes in total, of which the network correctly detected 19,488, with just 478 errors, and a total assertiveness of 97.61%. In the measurement of horizontal deviations, the method yielded a maximum error of 7.28∆, falling short of the literature-defined error limit of 8∆. In addition, when evaluating the result in the CV-I Dataset, the method achieves an accuracy equal to 100% e (95%CI = [0.77; 1]), 100% de specificity with (95%CI = [0.66; 1]) and 100% of sensitivity with (95%CI = [0.48; 1]) for horizontal strabismus. For the CV-II Dataset, the method obtained a sensitivity of 66.67% with (CI 95% [0.64; 0.99]), specificity of 100% with (CI 95% [0.93; 1]) and accuracy of 99.33% with (CI 95% [0,92; 1]) when considering only the horizontal measurement. Finally, when considering the horizontal and vertical measurements together, the method obtained a sensitivity of 66.67% with (CI 95% [0.63; 0.99]), specificity of 94.44% with (CI 95% [0.84; 0.98]) and accuracy of 93.89% with (CI 95% [0.84; 0.98]).