Aprendizado profundo para a detecção de pupilas dilatadas ou contraídas

Abstract

Pupillometry is the measurement of pupillary diameter used in some medical procedures to evaluate reactions of pupillary dilatation and constriction. Those reactions can be involuntary when it is caused by the Central Nervous System (CNS), or provoked by the light reflections. The iris and the pupil compose the ocular globe that is responsible for forming human vision. The iris is the biggest circumference with pigmented textures, which allows the formation of the color of the eyes. The pupil is the smallest circumference and it is in the internal region of the iris; the pupil is characterized by which allows the entry of light to form the vision. This region presents two common responses (states) through externals (reflections of light) and internals stimulus (Central Nervous System). The dilatation state occurs when there is an extension of the pupil size, in a constriction state. On the other hand, its size reduces. Understanding these pupillary reactions has been coming common among the researchers of neuroscience and cognitive psychology since it allows the identification of neurologic disorders. Therefore, pupillometry has been becoming a popular strategy in clinic pre-operatory processes, and in the identification of neurologic disorders in some individuals. Considering this context, this work aims to propose a computational method able to detect and measure pupil size, based on processing digital image techniques and machine learning to assist cognitive psychologists and neuroscience researchers to understand and identify neurologic diseases through pupillary reactions. The method of planning a multitasking neural network architecture with the inclusion of attention mechanisms, called At-Unet, to segment the iris and pupil region with the intention of obtaining the pupil diameter, and calculate the dilation factor that defines the state of the pupil in dilated or contracted. This method achieved 97.17% of Dice coefficient, from the cross experiment, so the type of pupil state estimated has an average error of the dilation factor of 0.0167.