CytoNet

Contrastive learning normally builds its training signal from augmentations: rotate, crop and recolour an image, and call the result the same thing. Cell-body-stained sections of human cortex break that assumption in both directions. The transformations that preserve semantics in natural photographs can destroy cytoarchitecture — the regionally specific arrangement of neurons into layers and columns that defines a cortical area — while the features an augmentation leaves untouched are often exactly the confounds, such as blood vessels and cortical folding geometry, that a mapping model should ignore.

CytoNet's answer is to take the similarity signal from anatomy instead of from pixels. The cortex is spatially continuous, so two image patches sampled a short distance apart along the cortical sheet almost certainly share an architecture, and two patches in different lobes almost certainly do not. Its SpatialNCE objective co-registers every section into the MNI Colin 27 reference space, converts pairwise 3D distances between sampled patches into continuous similarity weights through a radial basis function kernel, and trains the encoder to reproduce those weights. The coordinates define the loss but are never given to the network, so the model must infer spatially consistent cytoarchitecture from image content alone.

The model was developed at the Institute of Neuroscience and Medicine (INM-1) and Helmholtz AI of Forschungszentrum Jülich, with collaborators at Helmholtz Munich, University Hospital Düsseldorf and the University of Koblenz, and posted as a preprint in October 2025.

#Key Features

• Distance-weighted contrastive objective: SpatialNCE assigns continuous weights to all non-identical pairs in a batch from their anatomical separation, rather than picking discrete positive neighbours, and needs no manual annotation or augmentation-based positives.
• Evaluation on brains withheld from pretraining: five of the fourteen brains were excluded from self-supervised training and reserved for feature-space analysis and downstream testing, separating representation quality from memorisation of a specific specimen.
• Label efficiency: frozen features reach usable cortical layer segmentation from 1% of the annotated data, which matters because expert cytoarchitectonic annotation is the scarce resource in this field.
• Readouts beyond classification: the same feature vectors linearly predict cortical thickness, curvature, layer thicknesses and cell densities, and decode functional network identity.
• Published model zoo: weights, training configurations and precomputed feature embeddings for six variants are distributed openly.

#Technical Details

CytoNet is a family of encoders rather than a single network: modified ResNet18 and ResNet50 backbones with an enlarged input stem, and hybrids that append a ViT-B transformer to the ResNet feature map. The headline variant, CytoNet-R50-ViT (1M), has a 110.95M-parameter encoder. Pretraining used one million patches of 2,048 px at 2 µm/px — roughly a 4 mm field of view, enough to span the full cortical ribbon — drawn from over 4,000 sections across nine postmortem brains, for 150 epochs with LARS-augmented SGD at batch size 2,048 and an RBF bandwidth of σ = 10 mm. A run processed up to 4 TB per epoch and 600 TB in total on 64 A100 GPUs or 32 GH200 superchips.

On linear probing over 113 Julich Brain Atlas areas, CytoNet-R50-ViT (1M) reached macro-F1 0.72 on brains seen in pretraining, 0.38 on a transfer brain and 0.30 on unseen brains, against 0.28 for the strongest supervised contrastive baseline and 0.12 for the strongest pathology foundation model tested. Cortical layer segmentation under linear probing reached macro-F1 0.63 with 1% of the training data and saturated near 0.78, approaching UNI and Prov-GigaPath at full data and exceeding Virchow2, DINOv2 and MicroSAM. Finetuned on the independently acquired Allen Adult Human Brain Atlas, it reached macro-F1 0.65.

#Applications

The intended users are brain mappers working at histological scale. CytoNet supplies a single representation for classifying cytoarchitectonic areas across whole brains, segmenting cortical layers from sparse annotation, quantifying microarchitectural variation, and clustering features to propose subdivisions in regions the atlas has not yet charted, where it beat Virchow2 and BigBrain intensity profiles in all eight tested frontal-pole and Broca settings. Decoding functional parcellations from cytoarchitecture alone raised Yeo7 macro-F1 from 0.40 for a coordinate-only baseline to 0.63, offering a route to test structure–function correspondence in postmortem tissue.

#Impact

CytoNet moves cortical mapping from bespoke per-area classifiers toward a reusable pretrained representation, and shows spatial continuity is a viable self-supervision signal where augmentation-based contrastive learning fails. Its errors are anatomically plausible: roughly 80% of misclassifications on seen and transfer brains fall on directly adjacent atlas areas, concentrating at the borders where human experts also disagree. Cluster-selection criteria did not converge on a subdivision count, so unsupervised area discovery remains unsolved, and accuracy on brains never seen in pretraining trails that on seen ones. The work is a preprint awaiting peer review. Source code is Apache-2.0 and weights are public, but the histological image corpus is available only on request while curation and hosting infrastructure are completed.