Research groups at KU Leuven, Belgium, led by Prof. Dr. Wouter Peelaerts and Prof. Dr. Veerle Baekelandt have developed an AI-driven automated analysis pipeline with Aiforia® Create to perform continuous, high-throughput, and reproducible brain-wide histological quantification in mouse models of synucleinopathies(1). By replacing time-intensive stereology with multi-layer convolutional neural networks (CNNs) that analyze coronal tissue sections, the team analyzed up to 33 brain regions in hours rather than weeks, uncovering biological insights that conventional targeted sampling misses.
Nervous system disorders are now the primary driver of disease burden globally, with neurodegenerative disease on the track of becoming the second leading cause of death by 2040 as populations age(2,3). A critical subset of these conditions are synucleinopathies – neurodegenerative disorders characterized by the abnormal accumulation and aggregation of misfolded α-synuclein (α-syn) protein, including Parkinson’s disease, dementia with Lewy bodies, and multiple system atrophy (MSA). With an aging global population, neurodegenerative diseases are projected to become the second leading cause of death by 2040. To uncover the mechanism driving these disorders and speed the development of therapies, preclinical models have become central to finding solutions.
Read also >> Case study: AI image analysis reveals novel findings in Lewy Body Disease
Histological quantification remains the primary endpoint in preclinical neurodegenerative disease research for assessing treatment or evaluating disease progression. For decades, stereology has served as the gold standard for counting cells in brain sections and measuring pathological burden in tissue samples(4). However, manual stereology evaluation requires weeks of tedious effort by highly trained operators to process a single cohort, thus creating a substantial bottleneck in drug discovery and disease modeling.
"Histology is still the main readout in preclinical models of neurodegeneration, but quantifying it is slow, labour-intensive, and depends on who does the scoring. Stereology is the current standard. It can take weeks for a single cohort and only samples selected regions. We wanted to build an automated pipeline that matches stereology in accuracy, works across the whole mouse brain, removes inter-operator bias and takes a fraction of the time." — Prof. Dr. Wouter Peelaerts
Beyond the intensive resource demands, conventional microscopy and manual stereology face significant technical boundaries. Variations in scoring thresholds and operator experience introduce inter-laboratory discrepancies, making cross-study comparisons difficult. Most critically, because counting is performed manually and requires tens or hundreds of hours of work, researchers are constrained to predefined, smaller sets of regions of interest, such as the substantia nigra or the immediate injection site. This localized sampling unintentionally ignores distant brain regions, ruling out the broader anatomical context and unexpected histological observations.
Figure 1. Automated detection and spatial density mapping of dopaminergic neurons.
Panels illustrate automated, AI-driven cell segmentation and tissue-wide spatial density analysis of tyrosine hydroxylase (TH+), the primary marker used to assess dopaminergic cell loss in Parkinson’s disease models.
To address the quantification bottleneck in manual stereology, Dr. Anna Barber Janer, who led the work, and colleagues used Aiforia® Create to train five custom AI models tailored for PD and MSA research. Together, the AI models quantify i) dopaminergic neurons in the substantia nigra, ii) dopaminergic axonal loss in the striatum, iii) brain-wide neuron counts using the neuronal marker NeuN, iv) widespread microglial morphology and inflammatory activity with Iba1, and v) the accumulation of phosphorylated Ser129-α-synuclein across multiple brain regions.
The AI models were built using a nested, multi-layer architecture designed to mimic and automate a pathologist's analytical workflow, while significantly enhancing scalability and reproducibility. The first layer identifies the overall tissue section on the whole-slide image. The second layer automatically delineates anatomical structures across up to 33 distinct brain regions, eliminating the need to manually outline regions of interest. Finally, the third layer quantifies the target cells or protein aggregates within each identified brain region.
|
Figure 2. Automated hierarchical analysis and quantification of dopaminergic neurons using tyrosine hydroxylase (TH+) staining in coronal brain sections. Panels illustrate the multi-layer AI analysis workflow from whole-tissue segmentation and anatomical region mapping to high-resolution single-cell detection and profiling. |
Figure 3. Automated hierarchical analysis and anatomical mapping of microglial distribution. Ionized calcium-binding adaptor molecule 1 (Iba1) staining in whole-brain sections used to monitor neuroinflammatory responses. Panels illustrate the AI analysis workflow from whole-tissue segmentation to high-resolution profiling of microglial morphology. |
Aiforia® Create requires no programming expertise, allowing the team to train all five models entirely through visual annotations. In addition, the underlying Aiforia® Platform's cloud-based workspace enabled researchers to pool whole-slide images across different cohorts and laboratories, improving the robustness and generalizability of the trained algorithms.
"The platform let us build nested, multi-layer models. One layer finds the tissue, the next finds brain regions, and the last finds the cells or pathology of interest. Each layer only analyses what the layer above it passes on, and that is what made automatic, region-by-region analysis of the whole brain possible. Models are trained by annotating images, not by writing code, and applying a trained model needs no AI expertise. On the people side, much of our progress came from Aiforia's one-on-one support, which helped us design the pipelines we needed." — Prof. Dr. Wouter Peelaerts
When evaluated against annotations from three blinded expert validators, the AI models achieved F1-scores above 80% across most layers, mirroring the level of concordance among human experts. Neuronal loss estimates in the substantia nigra correlated closely with stereological benchmarks while demonstrating lower variance.
Crucially, the automated pipeline shortened the timeline for a full cohort analysis of 16 animals from weeks to just a few hours. This dramatic gain in throughput allowed the team to scan the entire mouse brain rather than relying on sparse sampling.
As a direct result of whole-brain profiling, the research group observed marked neuronal loss and microglial activation in distant brain regions far beyond the primary injection sites. In standard stereology setups, these pathological changes would have gone unnoticed simply because researchers lacked the time to analyze areas outside the predefined target region.
Figure 4. Automated morphology-based detection and classification of pathological protein aggregates.
Panels illustrate the original histological section (left) alongside AI-driven instance segmentation (right), categorizing distinct aggregate morphologies such as neuritic threads (blue) and somatic inclusions (magenta).
Beyond internal research gains, automated AI-driven image analysis addresses the field's pressing need for reproducible, continuous quantitative endpoints. By replacing subjective, semi-quantitative scoring with cell-level instance segmentation, researchers can capture subtle morphological shifts, such as changes in microglial ramification that signal inflammatory reactivity or the subcellular transition of α-synuclein from neuritic to somatic inclusions.
"Sharing is the part I find most exciting. If labs working with the same disease models use the same validated tools, their results become directly comparable, and the whole field moves faster. Brain-wide histological quantification at the level of single cells no longer has to be a bottleneck." — Prof. Dr. Wouter Peelaerts
The developed AI models have already supported peer-reviewed publications, including studies on chronic peripheral infections exacerbating synuclein pathology by Mercado et al. (2024) in the Journal of Parkinson's Disease and neuroinflammatory mechanisms in Alzheimer's disease by Dujardin et al. (2026) in Molecular Neurodegeneration Advances(5,6). Because trained AI models and Python post-processing scripts are easily shared across laboratories, external research teams can deploy identical evaluation criteria without setting up specialized stereology equipment or machine-learning infrastructure.
1) Barber-Janer, A., Van Acker, E., Vonck, E., Plessers, D., Rosada, F., Van den Haute, C., Baekelandt, V., & Peelaerts, W. (2025). Development of convolutional neural networks for automated brain-wide histopathological analysis in mouse models of synucleinopathies. NPJ Parkinson's Disease, 11(1), 317, https://doi.org/10.1038/s41531-025-01170-1
2) Gammon, K. (2014). Neurodegenerative disease: Brain windfall. Nature, 515(7526), 299–300. https://doi.org/10.1038/nj7526-299a
3) GBD 2021 Nervous System Disorders Collaborators (2024). Global, regional, and national burden of disorders affecting the nervous system, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. The Lancet. Neurology, 23(4), 344–381. https://doi.org/10.1016/S1474-4422(24)00038-3
4) Schmitz, C., Eastwood, B. S., Tappan, S. J., Glaser, J. R., Peterson, D. A., & Hof, P. R. (2014). Current automated 3D cell detection methods are not a suitable replacement for manual stereologic cell counting. Frontiers in neuroanatomy, 8, 27. https://doi.org/10.3389/fnana.2014.00027
5) Mercado, G., Clabout, A. C., Howland, V., Arkin, E., Janer, A. B., Plessers, D., Steiner, J. A., Smith, W. W., Hannan, T., Brundin, P., & Peelaerts, W. (2024). Chronic urinary tract infections cause persistent microglial changes in a humanized α-synuclein mouse model. Journal of Parkinson's Disease, 14(8), 1559–1574. https://doi.org/10.1177/1877718X241289046
6) Dujardin, P., Foroozandeh, J., Verhaegen, R. et al. (2026). Glial-specific and inducible full body C3 deficiency does not affect amyloid pathology in the AppNL−G−F mouse model of Alzheimer’s disease. Molecular Neurodegeneration Advances, 2, 25. https://link.springer.com/article/10.1186/s44477-026-00027-y