The combination of image analysis and fluorescence superresolution microscopy methods allows for unprecedented insight into the organization of macromolecular assemblies in cells. Advances in deep learning-based object recognition enables the automated processing of large amounts of data, resulting in high accuracy through averaging. However, while the analysis of highly symmetric structures of constant size allows for a resolution approaching the dimensions of structural biology, deep learning methods are prone to different forms of bias. A biased recognition of structures may prohibit the development of readouts for processes that involve significant changes in size or shape of amorphous macromolecular complexes. What is required to overcome this problem is a detailed investigation of potential sources of bias and the rigorous testing of trained models using real or simulated data covering a wide dynamic range of possible results. Here we combine single molecule localization-based superresolution microscopy of septin ring structures with the training of several different deep learning models for a quantitative investigation of bias resulting from different training approaches and finally quantitative changes in septin ring structures. We find that trade-off exists between measurement accuracy and the dynamic range of recognized phenotypes. Using our trained models, we furthermore find that septin ring size can be explained by the number of subunits they are assembled from alone. Our work provides a new experimental system for the investigation of septin polymerization.