Background: Proteins are of extremely vital importance in the human body, and no movement or activity can be performed without proteins. Currently, microscopy imaging technologies developed rapidly are employed to observe proteins in various cells and tissues. In addition, due to the complex and crowded cellular environments as well as various types and sizes of proteins, a considerable number of protein images are generated every day and cannot be classified manually. Therefore, an automatic and accurate method should be designed to properly solve and analyse protein images with mixed patterns.
Results: In this paper, we first propose a novel customized architecture with adaptive concatenate pooling and "buffering" layers in the classifier part, which could make the networks more adaptive to training and testing datasets, and develop a novel hard sampler at the end of our network to effectively mine the samples from small classes. Furthermore, a new loss is presented to handle the label imbalance based on the effectiveness of samples. In addition, in our method, several novel and effective optimization strategies are adopted to solve the difficult training-time optimization problem and further increase the accuracy by post-processing.
Conclusion: Our methods outperformed the SOTA method of multi-labelled protein classification on the HPA dataset, GapNet-PL, by above 2% in the F1 score. Therefore, experimental results based on the test set split from the Human Protein Atlas dataset show that our methods have good performance in automatically classifying multi-class and multi-labelled high-throughput microscopy protein images.
Keywords: DNNs; High-throughput microscopy images; Label imbalance; Multi-class and multi-label; Protein pattern recognition.