Source-linked AI summary
AxonDeepSeg: automatic axon and myelin segmentation from microscopy data using convolutional neural networks
Aldo Zaimi, Maxime Wabartha, Victor Herman, Pierre-Louis Antonsanti, Christian Samuel Perone, Julien Cohen-Adad
TL;DR
Axon and myelin segmentation methods often depend on modality-specific image processing and user parameter tuning, while few public methods are available. AxonDeepSeg addresses this gap with an open-source CNN framework and ready-to-use SEM and TEM models, achieving high pixel-wise accuracy across species and supporting morphological measurements from a full rat spinal cord slice.
Problem
Existing axon and myelin segmentation methods often require modality-specific processing steps and parameter tuning, and few methods are publicly available.
Method
AxonDeepSeg uses a U-Net-inspired convolutional neural network with ready-to-use SEM and TEM models and a pipeline for training models on new labelled data.
Results
Pixel-wise accuracy ranged from 81% on human SEM to over 94% on mice TEM, with morphological metrics from a full rat spinal cord slice overall consistent with known anatomy.
Takeaways & Limitations
AxonDeepSeg provides open-source segmentation software that supports axon and myelin analysis across microscopy modalities, species, and acquisition conditions.
Takeaways & Limitations
The models were trained and tested on healthy tissues, so performance on demyelinated samples remains to be assessed.
Abstract
from arXiv · showhide
Segmentation of axon and myelin from microscopy images of the nervous system provides useful quantitative information about the tissue microstructure, such as axon density and myelin thickness. This could be used for instance to document cell morphometry across species, or to validate novel non-invasive quantitative magnetic resonance imaging techniques. Most currently-available segmentation algorithms are based on standard image processing and usually require multiple processing steps and/or parameter tuning by the user to adapt to different modalities. Moreover, only few methods are publicly available. We introduce AxonDeepSeg, an open-source software that performs axon and myelin segmentation of microscopic images using deep learning. AxonDeepSeg features: (i) a convolutional neural network architecture; (ii) an easy training procedure to generate new models based on manually-labelled data and (iii) two ready-to-use models trained from scanning electron microscopy (SEM) and transmission electron microscopy (TEM). Results show high pixel-wise accuracy across various species: 85% on rat SEM, 81% on human SEM, 95% on mice TEM and 84% on macaque TEM. Segmentation of a full rat spinal cord slice is computed and morphological metrics are extracted and compared against the literature. AxonDeepSeg is freely available at https://github.com/neuropoly/axondeepseg
Methods
The experiments used SEM and TEM microscopy datasets spanning multiple species and tissue samples, with separate training and testing distributions for each modality.
- SEM acquisition resolutions ranged from 0.05 to 0.18 µm, whereas TEM resolutions ranged from 0.002 to 0.009 µm.
- Rat, human, mice, and macaque protocols were conducted under approval from the relevant ethics committees or with informed consent where applicable.
- SEM experiments trained on rat spinal cord samples and tested on rat and human spinal cord samples.
- TEM experiments trained on mice brain samples and tested on mice and macaque brain samples.
Gold standard labelling
Gold-standard masks were constructed through manual or manually corrected segmentation procedures for SEM and TEM images, with three final classes: background, myelin, and axon.
- SEM labels were created by manually tracing myelin contours, filling enclosed axon regions, and correcting contours or false positives when necessary.
- TEM labels used intensity thresholding with manual correction for myelin, followed by filling the inner region to generate axon labels.
- All labels were cross-checked by at least two researchers and encoded as background=0, myelin=127, and axon=255.
- Example SEM and TEM masks contain axon, myelin, and background classes in 512×512-pixel image crops.
Pipeline overview
AxonDeepSeg uses a four-step pipeline that standardizes microscopy inputs, trains and evaluates a network, then predicts on new images through patch-wise processing and reconstruction.
- The pipeline consists of data preparation, learning, evaluation, and prediction.
- Data preparation: Images and labels are resampled to 0.1 µm per pixel for SEM or 0.01 µm for TEM, then divided into 512×512 patches.
- Data preparation: Patches are randomly split into approximately 70/30% training and validation sets, while full test images are randomly selected for evaluation.
- Prediction: The trained model analyzes resampled 512×512 patches from new images, whose outputs are stitched and resampled back to native resolution.
- Learning and evaluation: The network is trained on the prepared training/validation data and assessed on a separate test dataset before inference on new microscopy images.
Architecture of the network
The AxonDeepSeg architecture is U-Net-inspired, combining contracting convolutions for context with expanding up-convolutions that recover localization information.
- The network combines a contracting path of convolutions with an expanding path of up-convolutions, following the original U-Net design.
- Stride-2 convolutions reduce feature dimensionality, while corresponding up-convolutions recover localization information during expansion.
- Features from contracting and expanding paths are concatenated to merge contextual and localization information.
- The SEM network uses three convolutional layers per block, whereas the TEM network uses two.
Data augmentation strategy
AxonDeepSeg augments training patches with geometric, intensity, and deformation transformations to reduce overfitting and increase variability.
- Data augmentation strategy: Shifting, rotation, rescaling, flipping, blurring, and elastic deformation were applied randomly to input training patches.The strategy was designed to reduce overfitting and improve generalization.
- Data augmentation strategy: The augmentation strategy and its corresponding parameters are summarized in Table 2.
Inference procedure
Inference uses overlapping 512×512-pixel patches and crops each prediction to a smaller output region to reduce border effects.
- Inference procedure: During inference, 512×512-pixel patches overlap by d pixels so the full image is covered.The default overlap value was d = 25.
- Inference procedure: Predictions are computed on an orange inference square, but only the inner white square is output before the window shifts across the image.The window advances by the size of the white output square.
Hyperparameter optimization
Hyperparameters were selected through grid searches using validation accuracy and error, with architecture and training parameters optimized jointly in stages.
- Hyperparameter optimization: Number of layers, filters, and convolutional kernel size were jointly optimized with grid searches over validation accuracy and error.
- Hyperparameter optimization: Starting learning rate and batch-normalization momentum were jointly optimized because they affect convergence time and validation-metric stability.
- Hyperparameter optimization: Batch-normalization momentum and its decay period were subsequently optimized jointly.
Evaluation method
Evaluation combines pixel-level segmentation measures with object-level detection measures to assess axon and myelin segmentation and myelinated-fiber detection.
- Evaluation method: Dice values for axon and myelin and pixel-wise accuracy assess segmentation quality.Pixel-wise accuracy measures correctly classified axon, myelin, and background pixels across the test sample.
- Evaluation method: The Dice coefficient is computed separately for axon and myelin predictions against gold-standard masks.It compares binary-image overlap using the intersection and the sizes of both masks.
- Evaluation method: Sensitivity and precision assess whether axonal fibers are detected while avoiding false axonal fibers.These object-level measures use axon centroid positions and counts of true positives, false positives, and false negatives.
Data availability
Dataset availability is partial: some datasets are deposited in the White Matter Microscopy Database, while the remainder can be obtained from the corresponding author on reasonable request.
- Some datasets generated or analyzed in the study are available through the White Matter Microscopy Database.
- The remaining datasets are available from the corresponding author on reasonable request.
Results
AxonDeepSeg was evaluated across SEM and TEM samples from multiple species and applied to a complete rat spinal cord slice to extract morphometric metrics.
- Validation: Segmentation was evaluated on rat and human spinal-cord SEM samples and mouse and macaque brain TEM samples.Validation included axon Dice, myelin Dice, pixel-wise accuracy, sensitivity, and precision.
- Validation: The example segmentations generally agree with gold-standard masks across SEM and TEM images from varied species.Discrepancies are associated with ambiguous myelin structure, uneven myelin thickness, and atypical axon intensity.
- Full-slice application: A full rat spinal cord SEM slice was segmented, with a zoomed view illustrating performance across smaller and larger axons.The smaller and larger axon regions had mean diameters of approximately 1.75 µm and 2.5 µm, respectively.
- Full-slice application: Morphometric statistics were extracted after downsampling the segmentation masks to 50×50 µm2 resolution.Computed measures included axon diameter, axon density, axon volume fraction, myelin volume fraction, and g-ratio.
- Full-slice application: The extracted metrics were restricted to white-matter pixels and compared with references for rat spinal-cord white-matter tracts.Distribution maps included axon diameter, density, volume fractions, and g-ratio.
Discussion
The discussion reports strong cross-species segmentation performance, demonstrates morphometric agreement with rat spinal-cord anatomy, and identifies modality-specific training and healthy-tissue evaluation as important boundaries.
- Limitations and future work: SEM and TEM models were trained at different target resolutions, and preliminary joint training produced lower performance than modality-specific models.The SEM and TEM target resolutions were 0.1 µm per pixel and 0.01 µm per pixel, respectively.
- Performance: Rat SEM pixel-wise accuracy was 85–88%, human SEM accuracy was 81%, and mouse TEM accuracy exceeded 94%.TEM performance on mouse samples also had sensitivity and precision around 96%.
- Performance: The lowest observed pixel-wise accuracy was 81%, supporting generalization across species and contrast changes.The authors compare this result with a maximal 82% pixel-wise accuracy reported for related optical-microscopy work.
- Morphometric application: Morphometric results from a full rat spinal cord were consistent with known tract anatomy and literature-reported g-ratios of 0.5–0.75.The reported tract-level patterns included larger axons in the ventral spinothalamic tract and higher density with smaller axons in the corticospinal tract.
- Software: AxonDeepSeg runs on Linux and Mac OS X, supports CPU inference, and segmented the full rat slice in about five hours on a Mac laptop.The software is open source and documented through GitHub resources.
- Limitations and future work: The models were trained and tested on healthy tissue, leaving performance on demyelinated samples for future assessment.Demyelinated tissue may contain thinner and differently shaped myelin sheaths.