Lysosomes Computational Labeling Method Based on Feature Map Slice of Deeplabv3+

According to the characteristics of the lysosomes in cell bright field images, a feature map slice of deeplabv3+ module is designed for lysosomes computational labeling. Firstly, as the features of the lysosomes target are rich in local details and small in size, a low-level feature map slice module is introduced to deeplabv3+ decoder module to enlarge the small lysosomes target in the local area, and improve the ability of labeling small lysosomes and representing detail information. Secondly, as the lysosomes area is small proportion in cell bright field images, the cross-entropy loss function is improved, which solves the imbalance between positive and negative samples in training process. Experiment on lysosomes dataset show that, the proposed method gets the recall rate 76.6%, precision 74.2% and F1-score 75.4%, which can extract the lysosome targets in cell bright field images completely and accurately.