Redes neurais para classificação de testes de eletroforese de imunofixação

Abstract

Annually, a single laboratory can perform thousands of immunofixation tests. The identification of the presence or absence of proteins in the sample is usually done visually, requiring the expertise of a specialist. A correct diagnosis is essential, as potentially harmful conditions, such as monoclonal gammopathy, can be diagnosed through the test. This study proposes two methods based on Machine Learning (ML) models to classify immunofixation tests. The methods use a Convolutional Neural Network (CNN) and a regular dense network with a manually engineered feature vector. Additionally, a preprocessing method is proposed to divide the image into vertical strips, which act as regions of interest. This method uses the color intensity histogram of the images to segment the image. The models were validated using a new database containing challenging immunofixation electrophoresis images. The results obtained with the VGGNet network were an F1-Score of 0.9088, precision of 0.9123, and recall of 0.9060. On the other hand, the results of the regular neural network were an F1-Score of 0.9625, precision of 0.9628, and recall of 0.9621. Both ML models demonstrated excellent performance in classifying immunofixation test images. Furthermore, we discuss why the regular neural network achieved similar or better metrics compared to the available state-of-the-art methods.