Extended Data Fig. 4: Performance validation of ERnet as a function of signal-to noise-ratio prevailing in ER image data.
(a) Pipeline to develop synthetic ground truth data to mimic ER structure as it would appear when imaged under the microscope including resolution loss through PSF blurring, and addition of image noise affecting the Signal to Noise ratio (SNR). A network was generated from points distributed randomly over a given field of view. Triangulation and tessellation were then used before cubic spine interpolation to obtain a network mimicking features of a fully connected tubular ER system. We added Gaussian noise and blurred images with a PSF kernel to produce ground truth data as obtained by our microscope setup. (b) Representative synthetic ER networks were generated at different noise levels. Ground truth images (left column) were processed, and Gaussian noise superimposed (middle column). Images were then analysed by ERnet to produce the skeleton map displaying the connectivity (right column), including nodes and edges. The output could then be directly compared with the results from ground truth data. (c) Degradation of network metrics as a function of decreasing image signal to noise ratio. The top panel compares the number of nodes obtained from the noisy data with that of the ground truth data (1 = fully matched, 0 = no match) as a function of noise level. The bottom panel shows how the number of three-way junctions (node of degree three) identified by ERnet decreases as noise increases. In the ground truth image (Noise level 0) ca 90% of all nodes constitute three-way junctions. For image signal to noise-ratios exceeding 5, ERnet reproduced the network topology to within 90% of the ground truth. N = 5 synthetic images per noise level. Data are presented as mean ± SEM. The algorithms and full instructions to generate synthetic ER networks mimicking microscopy data are available on GitHub and see Source Data Extended Data Fig. 4c. (https://github.com/charlesnchr/ERnet-v2).
