Assessing Knowledge Distillation of a Multi-Emitter Localizing Neural Network for Applications in Stochastic Optical Reconstruction Microscopy

Abstract

BackgroundSuper Resolution Microscopy (SRM) is a powerful method in quantitative bioscience that allows interrogation of nanoscale details. These methods require extensive imaging times on the microscope resulting in data sets on the scale of gigabytes. In order to reduce imaging times, the concentration of emitters can be increased, however that results in overlapping emitters rendering the isolation of single emitters extremely difficult. Statistical methods have been developed to deconvolute overlapping emitters, however they require parameter optimization and user expertise. Recently, Machine Learning (ML) has been developed to automate this analysis but often require larger Convolutional Neural Networks (CNN). While powerful, such models require compute and storage that would make pushing these models to compute limited devices difficult. To address this, we investigate if the dense multi-emitter localization capacity of a larger model, Deep Residual Stochastic Optical Reconstruction Microscopy (DRL-STORM), can be transferred to a smaller model, Super Resolution Convolutional Neural Network (SRCNN). ResultsKnowledge transfer from DRL-STORM to SRCNN did not result in an improvement of multi-emitter localization performance of SRCNN. Hint Learning (HL) was performed to facilitate knowledge transfer in a more deliberate manner. SRCNN demonstrated a limited capacity to learn an intermediate representation of the input image in the same manner as DRL-STORM and resultantly did not perform any better in its task. ConclusionsKnowledge transfer was not successful between DRL-STORM and SRCNN, but evidence suggests that it is possible and may require another model besides SRCNN. A future work will investigate if hyper-parameter optimization results in greater knowledge distillation between DRL-STORM and SRCNN. SRCNN may not be ideal for multi-emitter localizations, but it can still prove effective for SRM data analysis at emitter concentrations typical of SRM experiments and is uniquely suited as a neural network model that can be deployed in compute limited settings.