Classifying subcellular protein patterns in human cells

CLASSIFYING SUBCELLULAR PROTEIN PATTERNS IN HUMAN CELLS ABSTRACT Image recognition and object detection is a popular topic in the scientific field and industry today. There are many practical applications of classifying contents, especially through pictures. In the last decade, there have been important developments in this regard. In particular, machine learning has become the number one method in image recognition, as in other scientific fields, with the incredible increase in the amount of data available, computer hardware becoming more powerful and the many algorithms proposed. After the Artificial Neural Network, which was relatively insufficient for image recognition, the success of machine learning systems has increased with the invention of Convolutional Neural Networks. In this study, it was tried to find the types of proteins in the cell pictures. Due to the size of the data, an input pipeline has been developed for reading, training and evaluation processes. Then, three different models were designed, inspired by network architectures that have been successful in the past. Evaluating the results of these models, it has been tried to increase the amount of success by fine-tuning the most successful model. It is found from the loss and macro-F1 scores that the first model, which was designed with the inspiration of VGG, delivers the best results. In addition, it has been observed that the use of the dropout layer in cases involving overfitting generally works and increases the success of the system. Among the derivatives of the Gradient Descent algorithm, it was understood that the most successful for this dataset was Adam Optimizer, others did not give the desired result with default parameters. Keywords: Machine learning, image recognition, convolutional neural network, protein types in human cells, multi-class multi-label classification.