Improving Label-Free Nanoparticle Detection Accuracy in Interferometric Scattering Microscopy (iSCAT) Images Using a Mask R-CNN Machine Learning Framework

Interferometric Scattering Microscopy (iSCAT) has emerged as powerful technique for imaging nano-sized objects (i.e. proteins, nanoparticles, and viruses) with excellent spatiotemporal resolution without fluorescent labels. However, contrast in iSCAT is based on scattering, and image processing is required to gain meaningful information from experiments. The most essential step in processing, image segmentation, is complicated by strong scattering from features in the background of samples and remains a challenge. We improve upon traditional segmentation algorithms used in iSCAT processing by leveraging the power of instance image segmentation with machine learning using the Mask RCNN architecture. In a novel approach, we create a training dataset by superimposing point spread functions on background images from experiments collected on our iSCAT instrument. We then use transfer learning to train a Mask RCNN network to detect nanoparticles, reducing the time and resources needed for training. Improved performance is demonstrated via processing of a model nanoparticle adsorption experiment. Our processing workflow is not specific to iSCAT imaging, and we anticipate this methodology will improve image analysis in microscopy and single particle tracking research areas in general.