Lukas Schönenberger, Laurin Egli, Dimitrios Gkotsoulias, Christine Stadelmann, Cristina Granziera
HIDANet trained on 5,835 collected hatchery images reached a test accuracy of 98.44% with color inputs and 96.89% with grayscale inputs, with a macro-averaged F1-score of 0.99, which confirms strong morphological learning that is independent of chromatic background signals. An end-to-end workflow combining CLAHE image enhancement, Gaussian adaptive thresholding, and skeleton-based morphometric filtering offers automated larva counting, area and length measurements of the population, and density categorized into three levels (Low, Medium, and High).
Reliable automatic quantification of axon density in histologic sections is important for both research and clinical neuropathology applications - however, it remains a challenge for silver-impregnation stains, such as Bielschowsky. Classical color-deconvolution approaches perform poorly on silver stains, and manual annotation of densely packed axons is time-consuming and subject to inter-rater variability. We developed an automated axon segmentation pipeline in Bielschowsky silver-stained histological brain sections based on an attention U-Net architecture with attention-gated skip connections, trained using Focal Tversky Loss to handle class imbalance and thin structure recovery. Ground truth was manually annotated on 33 image tiles (covering over 25 million pixels) derived from 26 whole-slide images from varying brain regions of four multiple sclerosis patients. Slides were prepared by different laboratory technicians at different time points to capture realistic staining variability. Model performance was evaluated on a held-out test set of unseen tiles. On the test set (eight held-out tiles) the model achieved a mean pixel-wise F1/Dice of 0.717 and an intersection over union (IoU) of 0.584. A dedicated inter-rater experiment on two representative tiles yielded rater-to-rater F1 scores of 0.637 and 0.671, while model-to-rater agreement (F1: 0.681-0.736) met or exceeded that human ceiling. Generated whole-slide density and orientation visualizations accurately reflected regional axon distributions and provided quantitative readouts suitable for downstream analysis. Our attention U-Net provides an open-source solution for axon segmentation from Bielschowsky-stained sections, reducing manual effort and enabling reproducible, slide-level quantitative metrics. This tool can facilitate studies of axonal pathology across research and clinical settings.