Exploring different Convolutional Neural Networks architectures to identify cells in spheroids

• Autor(a) principal: SANTIAGO, A.G.
• Data de Publicação: 2020
• Outros Autores: CAMPOS, C.S., MACEDO, M.M.G., DAGUANO, J.K.M.B., DERNOWSEK, J.A., RODAS, A.C.D., BRAZILIAN CONGRESS IN BIOMEDICAL ENGINEERING, 27th
• Tipo de documento: Artigo de conferência
• Título da fonte: Repositório Institucional do IPEN
• Texto Completo: http://200.136.52.105/handle/123456789/31684
• topic: animal cells; spheroids; neural networks; cell cultures; fibroblasts; image processing

The cultivation of cells in 3D has gained more interest in research once 3D architecture can be closer to full cell physiological functionality. The cultivation of the cells in a spheroid format has shown very promising results, further for bioprinting developing so fast during the last decade. The interaction of spheroids and the matrix, or bioink, have proportionate new structures to be analyzed, specially if one would like to follow the whole system (spheroid and bioink) without fluorescent dyes. Trying to solve this image limitation, the aim of this paper is to present a study on different Convolutional Neural Networks (CNN) architectures employed to identify different structures in fibroblast NIH-3T3 spheroids. Three different architectures were considered: GoogleNet, ResNet18 and AlexNet, all implemented in Python 3.7 using the PyTorch Application Interface Programming (API). Given a spheroid image taken in a light microscope, four structures can be identified: the cell, the dead cell, the impurity/contamination and the background consisting of a gel in which the spheroid is immersed. All four CNN architectures were trained and evaluated with a dataset consisting of over 370 samples, split into a training set (??? 70%), a test set (??? 20%) and a validation set (??? 10%). Since our dataset has unbalanced classes, a data augmentation was applied in order to provide a comparable number of samples for all classes being considered.