Feature Reuse in CNN for Human Proteins Localization

Abstract

Human proteins localization plays a crucial role in determining cell activity, mind disease causes, and drug design. Nowadays, the use of microscopic fluorescence images in protein localization with computational methods has made protein maps of the human body closer to reality. Among the current methods, deep learning, especially convolutional neural networks, has successfully classified these images. In this study, we first propose a method for preprocessing the Human Protein Atlas (HPA) dataset and reducing data volumes up to 27 times without losing essential details. Then, proposing a novel convolutional neural networks (CNNs) architecture based on the two ideas of reusing summary features and designing block structures, we classify preprocessed images into 13 classes. Multi-labeling, large image sizes, and unbalanced data are a character of the data set’s challenges. Finally, reducing the required computing by about 50 percent less than state of the art and preserving plenty of storage space and needed memory, the image classification with macroF1-Score 0.789 excels among successful models.