cs.CVMar 26, 2026

LEMON: a foundation model for nuclear morphology in Computational Pathology

Authors: Loïc Chadoutaud, Alice Blondel, Hana Feki, Jacqueline Fontugne, Emmanuel Barillot, Thomas Walter

Organizations: Institut Curie, Paris, France · Mines Paris PSL, Centre for Computational Biology · INSERM U1331, Paris, France · Institut Curie, F-75005 Paris · Mines Paris PSL, Centre for Computational Biology (CBIO), F-75006 Paris · INSERM, U1331, F-75005 Paris · Institut Curie, U1353/UMR9029 IRIS, Equipe IMPACT, Paris, France · Department of Pathology, Université Paris-Saclay, UVSQ, Institut Curie, Saint-Cloud, France · Institut Curie, U1353/UMR9029 IRIS, Equipe IMPACT, F-75006 Paris

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Computational pathology relies on effective representation learning to support cancer research and precision medicine. Although self-supervised learning has driven major progress at the patch and whole-slide image levels, representation learning at the single-cell level remains comparatively underexplored, despite its importance for characterizing cell types and cellular phenotypes. We introduce LEMON (Learning Embeddings from Morphology Of Nuclei), a self-supervised foundation model for scalable single-cell image representation learning. Trained on millions of cell images from diverse tissues and cancer types, LEMON learns robust and versatile morphological representations that support large-scale single-cell analyses in pathology. We evaluate LEMON on five benchmark datasets across a range of prediction tasks and show that it provides strong performance, highlighting its potential as a new paradigm for cell-level computational pathology. Model weights are available at https://huggingface.co/aliceblondel/LEMON.

Figures & tables

Figure 1: Overview of the LEMON framework.
 MIDOG25NuCLSPanNukeHEST-BowelHEST-BreastHEST-Lung
 bal accbal accbal accpccpccpcc
Embeddings      
Natural images TL
Imagenet (ResNet-18)0.577 (0.008)0.693 (0.005)0.56 (0.004)0.114 (0.023)0.113 (0.023)0.083 (0.021)
Imagenet (ResNet-50)0.603 (0.012)0.711 (0.005)0.587 (0.005)0.12 (0.035)0.124 (0.023)0.091 (0.013)
Natural images FM
Table 1: Performance of pretraining strategy on downstream tasks. We report balanced accuracy for MIDOG25 (majority), NuCLS (super) and PanNuke (classification) and Pearson Correlation Coefficient for HEST (breast, lung, bowel). Values are mean (s.d.) over folds, best values are highlighted in bold.
TargetMIDOG25 (majority)
Metricbal accf1 (weighted)
Embeddings  
UNI-class (ViT-l/16)0.539 (0.007)0.834 (0.004)
UNI-center (ViT-l/16)0.645 (0.019)0.876 (0.007)
Virchow2-class (ViT-h/14)0.538 (0.005)0.833 (0.002)
Virchow2-center (ViT-h/14)0.673 (0.017)0.886 (0.006)
Table 2: Evaluation of representations from larger images ( 128×128128\times 128 pixel crop size) with patch level foundation models on the MIDOG25 dataset
Figure 2: Performance of pretraining strategies versus Floating point operations. Left: mean balanced accuracy across NuCLS (super, main, raw), MIDOG25 and PanNuke for classification. Right: mean Pearson Correlation Coefficient across breast, lung, and bowel datasets for regression. Marker size indicates model parameter count; color indicates pretraining category.
Figure 3: Performance of models vs. training data composition. Mean balanced accuracy (error bars represents standard errors) across MIDOG25 and the three NuCLS classification tasks. Top: Performance by training dataset size; models improve with larger datasets and level off near 1M images. Bottom: With the total number of images held constant, models perform better as diversity increases (more organs of origin and more slides).
 LEMON-MoCov3VoltaH-Optimus-0
 (ViT-s/8)(ResNet-18)(ViT-g/14)
Embedding Robustness
RMSE (↓)0.777 (0.150)0.879 (0.266)1.311 (0.076)
Cosine Similariy (↑)0.687 (0.116)0.579 (0.246)0.138 (0.099)
Neightborhood Overlap (↑)0.398 (0.067)0.297 (0.059)0.154 (0.014)
Biological tasks
Table 3: Robustness to Macenko stain normalization across 6 TCGA reference images. Embedding robustness metrics measure the shift in embedding space between original and stain-normalized cells. Biological tasks report balanced accuracy on stain-normalized embeddings. Best in bold.
Figure 4: Alignment between morphological embeddings and marker-gene expression in a bowel Xenium tissue section (HEST id: TENX147). Left: t-SNE projection of LEMON morphological embeddings. Right: normalized expression maps for marker genes, showing their spatial distribution within the morphology-derived manifold.
Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Table S1: NuCLS dataset with different annotation levels
ClassNumber of images
NMF10191
AMF1748
Appendix
Table S2: MIDOG25 dataset.
ClassNumber of images
NMF7955
AMF1160
Appendix
Table S3: MIDOG25 dataset with 128×128128\times 128 crop images
ClassNumber of images
039222
126108
217258
313457
41582
Appendix
Table S4: PanNuke dataset.
 MIDOG25NuCLSHEST-BowelHEST-BreastHEST-Lung
 bal accbal accpccpccpcc
Model     
DINOv2-Baseline (vits8)0.678 (0.011)0.681 (0.002)0.117 (0.014)0.119 (0.023)0.106 (0.015)
DINOv2-No local-to-global (vits8)0.673 (0.012)0.703 (0.003)0.131 (0.016)0.129 (0.021)0.110 (0.018)
DINOv2-No local-to-global + KDE (vits8)0.67 (0.011)0.767 (0.007)0.142 (0.024)0.143 (0.024)0.147 (0.02)
MoCov3 (vits8)0.741 (0.02)0.779 (0.005)0.16 (0.027)0.156 (0.021)0.128 (0.028)
Appendix
Table S5: Effect of architectural and framework adaptations on downstream tasks. We report balanced accuracy for NuCLS (super) and MIDOG25 (majority) and PCC for HEST datasets. Values are mean (s.d.), best values are highlighted in bold.
Figure S1: Examples of nuclei images with different augmentation strategies with three random augmentations from either the MoCo v3 augmentation or ours
 MIDOG25NuCLSHEST-BowelHEST-BreastHEST-Lung
 bal accbal accpccpccpcc
Augmentations     
a0 (MoCo)0.682 (0.013)0.771 (0.003)0.154 (0.03)0.148 (0.023)0.116 (0.027)
a10.715 (0.021)0.784 (0.003)0.161 (0.028)0.155 (0.025)0.133 (0.032)
a1+gray0.746 (0.015)0.779 (0.005)0.16 (0.027)0.156 (0.021)0.128 (0.028)
a1+gmm10.736 (0.018)0.768 (0.005)0.164 (0.019)0.162 (0.021)0.126 (0.029)
Appendix
Table S6: Effect of augmentation policy on downstream tasks. We report balanced accuracy for NuCLS (supervised) and MIDOG25 (majority), and Pearson Correlation Coefficient for HEST (breast, lung, bowel). Values are mean (s.d.) over folds, best values are highlighted in bold.
Figure S2: Visualization of Macenko stain augmentation. The first panel shows a subset of original MIDOG25 cell images, while the remaining panels display the same images after Macenko augmentation using six different reference stains.
Figure S3: Benchmark of performances under different Macenko staining augmentations. Balanced accuracy of three models (H-optimus-0, LEMON, and Volta) on the NuCLS (left) and MIDOG25 (right) datasets. For each dataset, results are shown for the original images, six Macenko-stained augmentation conditions, as well as the mean across these six augmentations.
Figure S4: Morphological map of learned cell representations. A t-distributed Stochastic Neighbor Embedding projection of 100,000 cells randomly sampled from a breast cancer whole-slide image (TCGA) is shown. A regular grid of cell images visualizes the embedding space, revealing a smooth continuum of latent morphologies. Five representative regions were annotated by an expert pathologist, corresponding to distinct cellular phenotypes: (1) epithelial cells, (2) neoplastic epithelial cells with large, irregular nuclei, (3) small lymphocytes, (4) plasma cells, and (5) fibroblasts.
TargetNuCLS (super)
Metricaucauprbal accf1 (macro)f1 (weighted)
Embeddings     
Imagenet (ResNet-18)0.876 (0.003)0.757 (0.005)0.693 (0.005)0.698 (0.006)0.74 (0.005)
Imagenet (ResNet-50)0.891 (0.004)0.777 (0.006)0.711 (0.005)0.717 (0.005)0.756 (0.004)
DINOv2 (ViT-s/14)0.905 (0.001)0.794 (0.002)0.732 (0.003)0.738 (0.004)0.775 (0.004)
DINOv2 (ViT-b/14)0.903 (0.003)0.792 (0.006)0.732 (0.003)0.738 (0.004)0.775 (0.003)
Appendix
Table S7: Evaluation on the NuCLS dataset - task super
TargetMIDOG25 (majority)
Metricaucauprbal accf1 (macro)f1 (weighted)
Embeddings     
Imagenet (ResNet-18)0.781 (0.01)0.948 (0.001)0.577 (0.008)0.597 (0.012)0.829 (0.004)
Imagenet (ResNet-50)0.805 (0.014)0.955 (0.003)0.603 (0.012)0.632 (0.016)0.84 (0.006)
DINOv2 (ViT-s/14)0.832 (0.011)0.961 (0.003)0.643 (0.004)0.677 (0.004)0.856 (0.002)
DINOv2 (ViT-b/14)0.842 (0.007)0.964 (0.001)0.653 (0.013)0.688 (0.013)0.86 (0.004)
Appendix
Table S8: Evaluation on the MIDOG25 dataset
TargetPanNuke (classification)
Metricaucauprbal accf1 (macro)f1 (weighted)
Embeddings     
Imagenet (ResNet-18)0.873 (0.002)0.589 (0.007)0.56 (0.004)0.583 (0.005)0.643 (0.003)
Imagenet (ResNet-50)0.885 (0.002)0.622 (0.005)0.587 (0.005)0.608 (0.005)0.663 (0.003)
DINOv2 (ViT-s/14)0.904 (0.002)0.661 (0.006)0.635 (0.009)0.653 (0.008)0.7 (0.003)
DINOv2 (ViT-b/14)0.901 (0.001)0.655 (0.005)0.627 (0.01)0.646 (0.008)0.691 (0.003)
Appendix
Table S9: Evaluation on the PanNuke dataset

Explore similar work

CardsList

  1. CellDETR: A Detection-Guided Framework for Scalable Cell Representation Learning from Histopathology Images

    Jun 28, 2026Shikang Zhang, Guojun Li, Yicong Mao +1Representation LearningDetection Transformer

    Zhejiang University of Technology · Tianjin University · Hangzhou Institute of Medicine, Chinese Academy of Sciences

  2. CellPath-Bench: A Multidimensional Benchmark for Whole-Slide Cellular Representations in Pathology Foundation Models

    Aug 21, 2026Bokai Zhao, Yiyang Zhang, Hanqing Chao +6Computational PathologyPathology Foundation Models

    School of Artificial Intelligence, University of Chinese Academy of Sciences. · Brainnetome Center, Institute of Automation, Chinese Academy of Sciences. · DAMO Academy, Alibaba Group. +1

  3. MOOZY: A Patient-First Foundation Model for Computational Pathology

    Mar 27, 2026Yousef Kotp, Vincent Quoc-Huy Trinh, Christopher Pal +1Computational PathologySelf-Supervised Pre-Training

    CSSE, Concordia University, Montréal, Canada · Mila – Québec AI Institute, Montréal, Canada · CHUM, Université de Montréal, Montréal, Canada +4