Membrane and vesicle structure detection in cryo-electron tomography based on deep learning

Cryo-electron tomography (cryo-ET) is a microscopy technique that enables the acquisition of 3D images of biological samples. Research in cell biology has shown that cellular processes are carried out by groups of macromolecules that interact in a crowded environment. In such an environment, where multiple biological macromolecules coexist and intertwine, semantic segmentation becomes even more challenging but crucial to understanding the structure and function of macromolecular complexes. However, manual semantic segmentation can be time-consuming, highly subjective, and prone to variability, which poses significant obstacles in studies dealing with large volumes of data. In contrast, automated algorithms such as Convolutional Neural Networks (CNNs) can process large-scale datasets with minimal human resources, thereby reducing the subjectivity associated with manual segmentation. In this work, we propose a convolutional neural network architecture that combines the features of U-Net, DeepLab, SegNet, Gated-SCNN, LSTM (Long Short-Term Memory), RNN (Recurrent Neural Network), and GAN (Generative Adversarial Network) architectures. This hybrid architecture effectively learns to identify different types of membranes and can replicate the behavior of a skilled human annotator. This system demonstrates a strong ability to segment various cellular membranes and vesicle structures.