Fig. 1: The analysis pipeline of ECLiPSE to quantify super-resolution microscopy data as point clouds applied to the validation data.
a, A schematic representation of the SMLM data acquisition process (t is time). b, Segmentation of the localizations into individual clusters (applied to a region of interest of the lysosome data using a maximum Voronoi area of 684 nm2 and a minimum of 25 localizations and subsequent filtering of lysosomes with an area less than 0.123 µm2 or greater than 0.479 µm2; green: low density and blue: high density). c, Feature extraction from segmented point cloud clusters generate descriptors including geometric, boundary, skeleton and so on. d, Example distributions of features that adequately or poorly separate the different classes in the validation data, as determined by automatic variable selection (28/67 descriptors that provide clear class separation). e, Data exploration using PCA with and without variable selection. f, Optimized classification results for the validation data (97.1 ± 0.1% accuracy), obtained by the random forest classifier (100 best models out of 1,000 generated models). g, Difference confusion matrices between ECLiPSE (logistic regression, no variable selection) and ASAP (10 nm rendering precision, 1.5 × 105 threshold, discriminant classifier). Left: validation data (96.9% versus 93.5% accuracy for ECLiPSE and ASAP, respectively). Right: tau aggregation data (92.9% versus 80.6% accuracy for ECLiPSE and ASAP, respectively). The blue values represent superior results for ECLiPSE (that is, positive diagonal values and negative off-diagonal values), whereas red values represent inferior results for ECLiPSE (that is, negative diagonal values and positive off-diagonal values).
