Redes neurais na classificação de imagens da membrana timpânica

Abstract

Otites are the main reason for preschoolers to visit a doctor. In the world, it is believed that approximately 740 million people per year will be affected by acute otitis media or suppurative chronic otitis media. Computers are increasingly helping health professionals to diagnose, prognosis and treat diseases. The interpretation of images in the medical field, mainly by deep learning neural networks, is an important tool in the aid of the diagnosis of diseases such as diabetic retinopathy. Radiology has been using neural networks to help identify tumors on CT scans, for example. Dermatology also uses these methods to help interpret images of skin blemishes to identify possible melanomas. These examples, like those in other areas of medicine, show the importance of neural networks in classifying images. In this work, images of the human external auditory canal were collected by otorhinolaryngologist, of which 133 were submitted to the Knowledge Discovery in Databases process. A Convolutional Neural Network model was constructed to classify these images into three classes: "without disease", "with disease" and "with cerumen". During training, the data was partitioned, and the average accuracy of all partitions was 81.62 %, the best result being 91.26 %. The final model was tested with 14 new images, presenting an accuracy of 90.47 %. (sic).