IVDAS: an interactive visual design and analysis system for image data symmetry detection of CNN models

Convolutional neural networks (CNNs) have achieved breakthrough performance in image recognition and classification. However, it is still a challenge for CNNs to classify images where the classification results strongly depend on the recognition of globally correlated features, for example, the identification of biomolecular structure symmetry in cryo-electron microscopy images. In this work, we improved the traditional CNNs model to solve the problem of cryo-electron microscopy image symmetry recognition and developed a visualization system to help biological researchers to design the best model. Our system consists of the model overview view, the model structure view, and the case view. The system supports a comprehensive analysis of the model’s recognition accuracy, the number of calculations, and the convolutional layers between different models. It also supports structural analysis of models and visual analysis of specific input images. Through the real cases from the biological researchers, we verify that our system can effectively build a model to identify symmetry. Graphic abstract