Figure 3
Overview of the NeMeAn data flow pipeline. AxonDeepSeg’s methods (A) were replicated to acquire semantic segmentations of myelin and axon from light microscopy images (resolution of 8 pixels/µm) stained with toluidine blue. To generate instance labels of myelinated fibers from semantic labels, the first data preparation step (B) clears erroneous semantic labels and then assigns myelin pixels to axons using the watershed algorithm. During the measurement collection step (C), each labeled myelinated fiber's morphometrics and window-based metrics are measured and collected into a “cell-wise” multivariate dataset. Separately in the same step (C), “pixel-wise” window-based metrics are collected for each possible window position in each sample. In the second data preparation step (D), multivariate measurements are split based on the sample type (control, sham, stim) and recombined into 2 datasets. Control and sham samples were used in the dataset for detecting changes related to surgical implantation of the cuff electrode. Sham and stimulated samples were used in the dataset for detecting changes related to stimulation level. These two datasets are then each split into training and validation sets. (E) A regression is fit to each of the surgery & stimulation training datasets and then evaluated on the data set aside for validation. Finally, both surgery & stimulation regressions are applied to all samples in the inference step (F).
