ECCV 2026

HASSL: Hierarchy-Aware Self-Supervised Learning
Framework for Single Cell Microscopy

Learning morphologically grounded, hierarchically consistent cell embeddings — without any labels.

Julius Riel1,2,3
Anuun Chinbat1
Hannes Leonhard1,3
Moritz Ladenburger1,3
Frederik Alexander1,3
Vishisht Choudhary1,3
Fabio Laredo4
Carsten Marr*5,6,7,8,9

Equal contribution    * Corresponding authors: {carsten.marr, amirhossein.kardoost}@helmholtz-munich.de

1TUM.ai, Munich   2Ludwig Maximilian University, Munich   3Technical University of Munich   4Institute of Stem Cell Research, Helmholtz Center Munich   5Institute of AI for Health, Helmholtz Munich   6Dept. of Medicine III, LMU Hospital   7Dept. of Physics, LMU   8German Cancer Consortium (DKTK)   9Munich Center for Machine Learning (MCML)
TUM.ai Helmholtz Munich LMU Munich Technical University of Munich MCML

Motivation

Hierarchy in cell imaging and HASSL embedding objective

Figure 1. Hierarchy in cell imaging (left) and the HASSL embedding objective (right). HASSL learns a more hierarchical embedding space, leading to improved morphological representation and tighter subclusters — separating OPCs from neurons even within the same imaging modality.

Different cell types can appear indistinguishable under the same imaging modality, while the same cell type can look markedly different across modalities. Standard SSL models pick up on modality as the dominant cue, forming coarse "superclusters" that obscure biologically meaningful substructure. HASSL pushes the model to form a latent space where cells are organised by morphology first.

Method

HASSL is a drop-in extension to DINO-style self-distillation. It adds two complementary objectives on top of the standard image-level DINO loss, with minimal compute overhead (+2.3% wall-clock, +10.2% VRAM over 100 epochs).

Component 1

Double-Teacher Distillation

A segmentation teacher is trained alongside the standard EMA image teacher. Zero-shot cell masks (CellposeSAM) provide structure-aware supervision that biases the student away from modality cues and toward cell shape, initiating the emergence of morphology-based subclusters via a convex combination of three cross-view distillation terms: image→image, seg→seg, and image→seg.

Component 2

Hierarchy-Aware Contrastive Loss

At each training step, HDBSCAN is run on the current batch embeddings to obtain a condensed cluster tree. For each anchor, stability-weighted (λ) positive and negative prototypes are mined at every level of the hierarchy. The resulting hinge-contrastive loss pulls each cell toward its ancestor cluster centroids while repelling sibling subtypes, compressing subclusters and widening ambiguous boundaries.

HASSL training pipeline overview

Figure 2. Overview of the HASSL training pipeline. (1) Zero-shot segmentation maps are generated per cell. (2) A student ViT is trained via double-teacher distillation from a global image teacher and a segmentation teacher. (3) HDBSCAN derives a cluster hierarchy from batch embeddings; stability-λ-weighted prototypes define positive and negative pairs at each hierarchical level. (4) The resulting latent space organises cells into superclusters with clearer morphology-driven subclusters (astrocytes, neurons, OPCs, dead cells).

Dataset

We curate a large-scale single-cell microscopy corpus spanning diverse imaging modalities, cell types, and biological contexts to train and evaluate hierarchy-aware representations.

2.3M
Single cell crops
20
Microscopy datasets
208
Cell classes
8
Imaging modalities

Results

We evaluate with a k-NN retrieval protocol on global embeddings across all test datasets, reporting Acc@K, Prec@K, mAP, and clustering agreement (NMI / AMI). All baselines are retrained on our split with the same resources for fairness.

Method Acc@1 Acc@3 Acc@5 Acc@9 mAP NMI ↑ AMI ↑
Cellpaint-DINO44.855.761.768.949.747.046.4
ChadaViT46.358.171.779.452.945.244.6
OpenPhenom45.357.163.971.950.840.339.5
scDINO50.461.968.075.355.443.943.3
HCSC*46.161.664.579.753.945.344.6
Baseline DINOv3*49.464.472.180.956.546.846.2
DINOv3 + DBSCAN49.161.668.175.655.247.246.6
DINOv3 + Unweighted HDBSCAN48.460.667.274.554.347.046.3
HASSL (w/o Double Teacher)50.966.674.382.858.147.646.9
HASSL (w/o HDBSCAN)50.667.675.182.957.447.246.2
HASSL (Ours)50.968.075.583.557.847.947.3

Table 1. Top-k retrieval and clustering agreement (%). * retrained on our subset for fairness. Stability-weighted HDBSCAN and Double-Teacher distillation provide complementary gains; their combination yields the strongest clustering agreement.

t-SNE latent space visualization

Figure 3. Latent space visualization on the test set (t-SNE, dim=2). In the internal stem cell data (orange), double-teacher distillation brings superclusters of different modalities but the same cell type closer, while HDBSCAN keeps subclusters compact and discernible.

Top-k retrieval on flat vs deep hierarchy datasets

Figure 4. Top-k retrieval on flat (depth=1, left) vs. deep hierarchy (depth>1, right) dataset subsets. HASSL improves by +6.3% at K=9 on multi-level hierarchy datasets without compromising performance on flat ones.

Downstream Classification

We train a frozen-backbone MLP classifier on our curated dataset and two held-out sets (Human Protein Atlas and the Allen Institute Drug Perturbation Dataset) to assess generalisation and biological relevance.

MethodAcc (%)F1 macroF1 weighted
HASSL (Ours)45.648.843.6
Baseline DINOv344.947.642.8
Cellpaint-DINO43.449.441.4
ChadaViT42.646.140.6
HCSC41.943.141.9
scDINO37.637.734.4
OpenPhenom34.728.631.7

Table 2. Curated dataset cell-type classification.

MethodAcc (%)F1 macroF1 weighted
HASSL (Ours)92.288.992.0
OpenPhenom85.775.884.2
Cellpaint-DINO84.271.881.7
Baseline DINOv381.977.181.8
ChadaViT79.272.578.3
HCSC64.456.263.5
scDINO65.053.863.4

Table 3. Drug perturbation identification (Allen Institute). HASSL leads by +7.8 F1weighted. F1 weighted averages per-class F1 scores weighted by class support (number of samples).

Citation

If you find this work useful, please consider citing:

@article{riel2026hassl, title = {HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy}, author = {Julius Riel and Vishwa Mohan Singh and Sai Anirudh Aryasomayajula and Anuun Chinbat and Hannes Leonhard and Moritz Ladenburger and Frederik Alexander and Vishisht Choudhary and Fabio Laredo and Giacomo Masserdotti and Thorben Prein and Carsten Marr and Amirhossein Kardoost}, journal = {arXiv preprint arXiv:2607.04353}, year = {2026}, url = {https://arxiv.org/abs/2607.04353}, }