Quality control in structured illumination-based super-resolution FRET imaging via machine learning

Abstract

Structured illumination-based super-resolution Förster resonance energy transfer microscopy (SISR-FRETM) has facilitated better observation of molecular behavior in living cells. However, SIM tends to produce artifacts in reconstruction, especially when the raw SIM inputs are of low signal-to-noise ratio (SNR) or out-of-focus, leading to erroneous signals in subsequent FRET. Current SIM quality evaluation metrics fail to utilize both SNR and out-of-focus features, making it challenging to classify unqualified raw data for FRET. Here, we propose an ensemble machine learning based SISR-FRETM quality control algorithm (SFQC) to evaluate the quality of SISR-FRETM raw data from the perspective of both SNR and focus quality. Specifically, SFQC extracts features with both SNR and focus quality metrics and combines them as feature vectors for machine learning models to train. To ensure high robustness of quality control, four different classifiers are trained and ensembled. In our experiment, SFQC is demonstrated to surpass all conventional SIM quality metrics on the F1-score up to 0.93 for the focus detection task and 0.95 for the SNR detection task, while also achieving the fastest processing time compared to other metrics. SFQC also provides options for researchers to generate focus error maps for error localization and masking for FRET results. Overall, by combining different quality metrics, we propose SFQC as an accurate, fast solution for selecting trust-worthy images of SR quantitative FRET imaging microscopy, which saves scientists from tedious human efforts on large scale microscopy image quality control works.