Non-destructive Analysis of Spheroids Using Artificial Intelligence: Preliminary Study of Convolutional Neural Networks Performance

With an increasing research of new 3D cell cultures, specially spheroids, Convolutional Neural Networks (CNNs) have proven their potential for different classification tasks in the analysis of these biological images, and choosing the most suitable CNN for these tasks is a challenge due to a variety of algorithms specially concerning features extraction. This paper presents a non-destructive spheroid cell classification using CNN’s and three different image datasets of fibroblast NIH-3T3 spheroids acquired in different culture conditions, each consisting of approximately 300 cell samples. The first two sets possess four possible cellular structures: living cells inside spheroids; living cells outside spheroids; dead cells; and background, while the third consists only of living cells inside the spheroid and background. Each dataset was sampled with a proportion of 70–20–10% for training-testing-validation. In order to improve the classification task and avoid overfitting, i.e., the lack of generalization of the network, several strategies were applied, such adding blurred images to the training dataset. The CNN architectures chosen were AlexNet, GoogleNet, MobileNetV3 Large, ResNeXt, SqueezeNet, Wide ResNet-50–2 and VGG16 with Batch Normalization, each presenting a different technology for the feature extraction layer and the classification layer as well. ResNeXt presented the highest F1-score with mean value of 0.9915 and was chosen as the best CNN for this application.