Open-set Recognition of Unseen Macromolecules in Cellular Electron Cryo-Tomograms by Soft Large Margin Centralized Cosine LossXuefeng Du (Xi'an Jiaotong University), Xiangrui Zeng (Carnegie Mellon University), Bo Zhou (Yale University), Alex Singh (Carnegie Mellon University), Min Xu (Carnegie Mellon University) Abstract
Cellular Electron Cryo-Tomography (CECT) is a 3D imaging tool that visualizes the structure and spatial organization of macromolecules at sub-molecular resolution in a near native state, allowing systematic analysis of seen and unseen macromolecules. Methods for high-throughput subtomogram classification on known macromolecules based on deep learning have been developed. However, the learned features guided by either the regular Softmax loss or traditional feature descriptors are not well applicable in the open-set recognition scenarios where the testing data and the training data have a different label space. In other words, the testing data contain novel structural classes unseen in the training data. In this paper, we propose a novel loss function for deep neural networks to extract discriminative features for unseen macromolecular structure recognition in CECT, called Soft Large Margin Centralized Cosine Loss (Soft LMCCL). Our Soft LMCCL projects 3D images into a normalized hypersphere that generates features with a large inter-class variance and a low intra-class variance, which can better generalize across data with different classes and in different datasets. Our experiments on CECT subtomogram recognition tasks using both simulation data and real data demonstrate that we are able to achieve significantly better verification accuracy and reliability compared to classic loss functions. In summary, our Soft LMCCL is a useful design in our detection task of unseen structures and is potentially useful in other similar open-set scenarios.