Classificação de imagens de ultrassonografia com redes neurais convolucionais em ambiente de alto desempenho

Abstract

Advances in the computerized analysis of medical images provide the emergence of methods and techniques that may assist in the early diagnosis of a number of diseases. Currently, these diagnoses are so due to visual inspection by human experts. This paper describes the development of a computational model, based on Convolutional Neural Network (CNN), that aims to assist the radiologist in the confirmation of a diagnosis by identifying masses in findings of ultrasonographic images of the liver. In general, the process of training a neural network requires a high processing power for dense data. When it comes to image processing, this issue is further enhanced, since the number of network inputs is proportional to the number of pixels of the image of interest, justifying the use of parallel processing. It is important to note that the medical imaging class does not allow lossy data storage and requires precission in its identification. The developed CNN model was able to perform the classification of these masses with an accuracy of 85% , for the set of analyzed images.