cs.CVMar 10, 2026

Deep Learning for BioImaging: What Are We Really Learning?

Authors: Ivan Svatko, Maxime Sanchez, Ihab Bendidi, Gilles Cottrell, Auguste Genovesio

Organizations: Université Paris Cité, IRD, Inserm, MERIT, F-75006, Paris, France · IBENS, École Normale Supérieure, Université PSL, Paris, France · Institut Curie, Université PSL, Paris, France · Iktos, Paris, France · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Representation learning has driven major advances in natural image analysis by enabling models to acquire high-level semantic features. In microscopy imaging, however, it remains unclear what current representation learning methods really learn. In this work, we conduct a systematic study of representation learning for the two most widely used and broadly available microscopy data types, representing critical scales in biology: cell culture and tissue imaging. We investigate whether, in contrast to natural images, existing models fail to consistently acquire high-level, biologically meaningful features. To this end, we introduce a set of simple yet revealing baselines on curated benchmarks, including untrained models and structural representations of cellular tissue. Our results show that, surprisingly, for a considerable subset of evaluation settings, the baselines are comparable to state-of-the-art methods, demonstrating that many commonly used benchmark metrics are insufficient to assess representation quality and often mask a lack of relevant high-level abstractions. In addition, we investigate how detailed comparisons with these baselines provide ways to interpret the strengths and weaknesses of models for further improvements. Together, our results suggest that progress in representation learning for microscopy requires not only stronger models, but also benchmarks that are more indicative of what is actually learned.

Figures & tables

Figure 1: Comparison of aggregated tissue patches (left) and 5-nearest-neighbor cell-graph images (right) . Both images are rendered at 224 ×\times 224 pixels. Aggregating tissue patches can expose larger-scale tissue organization in the graph rendering. Cell nuclei are revealed in a dark blue color via H&E staining.
Figure 2: Pretrained versus untrained representations on natural and microscopy images. Left : gene–gene retrieval performance on RxRx3-core (Recall@5 %\% ). Mean recall across all 5 literature datasets is reported. Right : kNN top-1 accuracy on ImageNet-1k across models.
Figure 3: Spearman ranking correlation of gene–gene rankings across model architectures. The heatmap displays pairwise Spearman correlations between gene–gene similarity rankings induced by different representations on RxRx3-core genes across 269,745 pairs. Models are hierarchically clustered to reveal functional groupings based on their representational alignment. Labels prefixed with [P] (blue) denote pretrained models, while [U] (red) indicates untrained/random initializations.
Figure 4: Layer-wise evolution of representation quality across domains. Average performance across architectures as a function of network depth for natural images (ImageNet-1k kNN accuracy) and cell-culture microscopy (RxRx3-core recall@5 %\% ). Average, minimal, and maximal scores are reported for each configuration.
Figure 5: Mean average precision (mAP) per compound on the JUMP-CP benchmark. Bar plots show mAP scores for eight positive-control compounds. Each bar corresponds to a model configuration, grouped by architecture family and colored accordingly. Solid bars denote pretrained models, hatched bars denote untrained models, and pixel-based baselines are included for reference. Error bars indicate variability across evaluation folds. The dashed horizontal line indicates the theoretical random baseline.
ModalityTrainingCCRCCPRADSKCMCOAD
Full ImageNo0.12 ±\pm 0.140.25 ±\pm 0.060.34 ±\pm 0.060.27 ±\pm 0.08
Full ImageYes, ood0.26 ±\pm 0.070.41 ±\pm 0.010.62 ±\pm 0.070.30 ±\pm 0.07
Full ImageYes, iid0.37 ±\pm 0.060.49 ±\pm 0.020.68 ±\pm 0.030.38 ±\pm 0.06
StructureYes, ood0.16 ±\pm 0.040.38 ±\pm 0.040.45 ±\pm 0.070.31 ±\pm 0.01
Table 1: Performance of the best-in-class model on a selection of HEST-1k-1NN datasets. Mean and standard deviation across cross-validation folds are reported. The best structure-based models were pretrained on natural images.
Figure 6: Gene-wise average test PCC across cross-validation folds. The best structure-based model is compared against H-Optimus-1 on selected datasets from HEST-1k-1NN. Each point represents one of the 50 dataset-specific target genes.
Figure 7: Layer-wise performance on HEST-1k-1NN. Comparison of performance of out-of-domain pretrained and untrained models and foundation models trained on biological modalities. For each architecture the global average across the benchmark and minimum/ maximum score across model configurations are reported.
ModelTrainingCCRCCPRADSKCMCOAD
Cell countNo0.11 ±\pm 0.070.37 ±\pm 0.050.40 ±\pm 0.060.30 ±\pm 0.07
GCNYes, synth. data0.04 (-0.02)0.38 (+0.04)0.45 (+0.08)0.31 (+0.04)
EGNNYes, synth. data0.04 (-0.01)0.36 (+0.005)0.22 (-0.05)0.28 (-0.02)
ResNet152Yes, IN1k0.11 (+0.04)0.34 (+0.09)0.37 (+0.23)0.24 (+0.14)
Table 2: Impact of pretraining on structure-only models on HEST-1k-1NN. We report average gene-wise PCC per dataset and the absolute improvement over the untrained version in parenthesis. Standard deviations for cell count are computed over the cross-validation folds.
Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Samples from 8 classes sampled for the mAP experiment.
Figure 9: Example of images from the 8 positive controls of JUMP-CP.
Figure 10: Outlier image detection on RxRx3-Core using raw image representations. Left: UMAP projections of embeddings produced by different encoders colored by cell count, highlighting an outlier cluster in red. Right: Representative images sampled from the identified outlier cluster.
TaskNumber of slidesNumber of patients (splits)Condition
IDC44Invasive ductal carcinoma (breast)
PRAD232Prostate adenocarcinoma
PAAD22Pancreatic adenocarcinoma
SCKM22skin cutaneous melanoma
COAD42colon adenocarcinoma
READ42rectal adenocarcinoma
Appendix
Table 3: Benchmark subsets of HEST-1k as defined in ( Jaume et al., 2024 ) . We keep the original structure of the benchmark unchanged.
Figure 11: We display neighbor counts for slides with different ST grids. The counts are given in the logarithmic scale.
Architecturetimm Model Nametimm WeightsModel StagesAggregation
Convolutional Neural Networks (CNNs)
ResNet18resnet18tv_in1klayer ∈\in {1, 2, 3, 4}Avg-pool
ResNet34resnet34tv_in1klayer ∈\in {1, 2, 3, 4}Avg-pool
ResNet50resnet50tv2_in1klayer ∈\in {1, 2, 3, 4}Avg-pool
ResNet152resnet152tv2_in1klayer ∈\in {1, 2, 3, 4}Avg-pool
Vision Transformers (ViTs)
Appendix
Table 4: Configurations of ResNet ( He et al., 2016 ) and ViT ( Dosovitskiy et al., 2021 ) used in our experiments.
Figure 12: Multiscale views in WSIs. (a) Comparison of the original patches from HEST-1k and HEST-1k-1NN and the corresponding cell graphs. (b) The source WSI from which the samples are taken. The slide overlays a square grid of ST spots and the aggregated patch corresponds to a 9-spot neighborhood.
ClassSamplingNoise and/ or densityCell count
0Noisy Grid, 10 voxelsσ(x,y)∝Id\sigma(x,y)\propto\text{Id}900
1Noisy Grid, 10 voxelsσ(x,y)∝SlopeLandscape(x,y;3,3)\sigma(x,y)\propto\text{SlopeLandscape}(x,y;3,3)900
2Noisy Grid, 10 voxelsλ(x,y)∝SlopeLandscape(x,y;3,3)\lambda(x,y)\propto\text{SlopeLandscape}(x,y;3,3) , σ(x,y)≡0.01\sigma(x,y)\equiv 0.01890
3Noisy Grid, 4 voxelsλ(x,y)∝StepLandscape(x,y;0.5,1)\lambda(x,y)\propto\text{StepLandscape}(x,y;0.5,1) , σ(x,y)≡0.01\sigma(x,y)\equiv 0.01848
4Noisy Grid, 10 voxelsλ(x,y)∝Discs(x,y;1,0.1,emboss,2)\lambda(x,y)\propto\text{Discs}(x,y;1,0.1,\text{emboss},2) , σ(x,y)≡0.01\sigma(x,y)\equiv 0.01964
5Noisy Grid, 10 voxelsλ(x,y)∝Discs(x,y;3,0.1,emboss,2)\lambda(x,y)\propto\text{Discs}(x,y;3,0.1,\text{emboss},2) , σ(x,y)≡0.01\sigma(x,y)\equiv 0.011028
Appendix
Table 5: A Summary of Synthetic Classes.
ModelAverage F1 ScoreWorst Class (F1-score)Best Class (F1-score)Invariance
ResNet18 (from scratch)0.87class 18 (0.38)classes 1,3,5,7,8,9,14,15,17,19,21 ( ≥\geq 0.99 )Learned, data augmentations
ViT-l/16 (linear probing)0.68class 22 (0.15)classes 7,9,12,15,17,20,21 ( ≥\geq 0.98)Learned, data augmentations
GCN (from scratch)0.72class 4 (0.15)classes 7,8,9,12,15,17,21 ( ≥\geq 0.98)Learned, data augmentations
E2GNN (from scratch)0.73class 4 (0.04)classes 1,5,12,17,18,19 (=1.0)Architectural, E( 22 )
Appendix
Table 6: Test scores on our 24-class synthetic dataset. Classes with F1-scores above the threshold of 0.98 are provided.
Figure 13: Illustration of synthetic graph data. Samples are designed to represent variations in local density while controlling for the node count. (a) Selected intensity functions that are used to scale density λ\lambda and spatial noise σ\sigma as listed in table 5 . (b) Test samples for each of 24 classes.
Figure 14: Cell count and slide separation in GNNs. UMAP plots of patch-level embeddings extracted with an untrained and trained GCN on LYMPH IDC-1NN.
Figure 15: Comparison between HEST-1k and HEST-1k-1NN. (a) Alignment of average PCC for a given dataset-model configuration pair. (b) Correlations between the performance of histology foundation models.
Figure 16: Experiments on the non-binned HEST-1k. Structure-based models remain competitive on a subset of genes across all ranges of PCC scores. (a) Gene-wise PCC for images of cell graphs and an in-domain foundation model H-optimus-1. (b) Gene-wise PCC for images of cell graphs and an out-of-domain DINOv2-g/14.
Figure 17: Possible contribution of cell count. Each point represents gene-wise PCC for images of cell graphs and a non-linear cell-counting baseline. (a) Binned HEST-1k-1NN. (b) Non-binnned HEST-1k.
  BenchmarkIDCPRADPAADSKCMCOADREADCCRCCLUNGLYMPH IDCAvg
ModalityTrainingModel          
1x1Hand-craftedCell count (PCA+ridge)0.32 ±\pm 0.040.26 ±\pm 0.020.32 ±\pm 0.040.26 ±\pm 0.050.26 ±\pm 0.0040.02 ±\pm 0.0010.04 ±\pm 0.070.42 ±\pm 0.0030.11 ±\pm 0.050.22 ±\pm 0.13
Cell count (XGBoost)0.31 ±\pm 0.050.26 ±\pm 0.020.34 ±\pm 0.080.30 ±\pm 0.070.26 ±\pm 0.010.05 ±\pm 0.030.03 ±\pm 0.060.44 ±\pm 0.010.11 ±\pm 0.070.23 ±\pm 0.12
1NNHand-craftedCell count (PCA+ridge)0.38 ±\pm 0.060.37 ±\pm 0.050.30 ±\pm 0.040.32 ±\pm 0.020.30 ±\pm 0.07-0.002 ±\pm 0.030.11 ±\pm 0.070.49 ±\pm 0.020.21 ±\pm 0.110.28 ±\pm 0.14
Cell count (XGBoost)0.37 ±\pm 0.080.36 ±\pm 0.050.34 ±\pm 0.070.40 ±\pm 0.060.30 ±\pm 0.070.03 ±\pm 0.020.04 ±\pm 0.090.49 ±\pm 0.020.21 ±\pm 0.160.28 ±\pm 0.15
Synthetic dataEGNN0.22 ±\pm 0.070.36 ±\pm 0.090.33 ±\pm 0.090.22 ±\pm 0.060.28 ±\pm 0.120.02 ±\pm 0.060.05 ±\pm 0.0680.19 ±\pm 0.090.17 ±\pm 0.090.20 ±\pm 0.12
Appendix
Table 7: Pearson Correlation Coefficient (PCC) across datasets for cell-counting and GNN baselines for HEST-1k and HEST-1k-1NN . The best and the second best score for HEST-1k-1NN are in bold and underlined respectively. Standard deviations across the cross-validation folds are reported.
  BenchmarkIDCPRADPAADSKCMCOADREADCCRCCLUNGLYMPH IDCAvg
ModalityTrainingModel          
1x1-imgFoundation ModelCONCH v10.536 ±\pm 0.0840.355 ±\pm 0.0100.446 ±\pm 0.0710.579 ±\pm 0.0500.253 ±\pm 0.0080.163 ±\pm 0.0490.218 ±\pm 0.0350.531 ±\pm 0.0110.251 ±\pm 0.0420.370 ±\pm 0.157
CONCH v1.50.544 ±\pm 0.0850.381 ±\pm 0.0100.457 ±\pm 0.0600.554 ±\pm 0.0360.279 ±\pm 0.0130.159 ±\pm 0.0660.218 ±\pm 0.0390.550 ±\pm 0.0070.270 ±\pm 0.0540.379 ±\pm 0.154
CTransPath0.511 ±\pm 0.0530.343 ±\pm 0.0460.436 ±\pm 0.0670.512 ±\pm 0.0810.228 ±\pm 0.0570.113 ±\pm 0.0770.228 ±\pm 0.0470.503 ±\pm 0.0400.235 ±\pm 0.0480.345 ±\pm 0.151
GigaPath0.553 ±\pm 0.0730.370 ±\pm 0.0210.474 ±\pm 0.0490.556 ±\pm 0.0670.291 ±\pm 0.0250.193 ±\pm 0.0670.241 ±\pm 0.0380.544 ±\pm 0.0350.252 ±\pm 0.0520.386 ±\pm 0.148
H-Optimus-00.598 ±\pm 0.0840.379 ±\pm 0.0020.492 ±\pm 0.0410.655 ±\pm 0.0580.302 ±\pm 0.0020.222 ±\pm 0.0480.274 ±\pm 0.0350.561 ±\pm 0.0280.260 ±\pm 0.0430.416 ±\pm 0.163
Appendix
Table 8: Pearson Correlation Coefficient (PCC) across datasets and foundation models for HEST-1k and HEST-1k-1NN . The best and the second best score per modality are in bold and underlined respectively. Standard deviations across the cross-validation folds are reported.
  BenchmarkIDCPRADPAADSKCMCOADREADCCRCCLUNGLYMPH IDCAvg
ModalityTrainingModel          
1NN-graphPretrainedResNet1520.340 ±\pm 0.0360.338 ±\pm 0.0170.312 ±\pm 0.0720.366 ±\pm 0.0910.243 ±\pm 0.0400.090 ±\pm 0.0580.111 ±\pm 0.0300.382 ±\pm 0.0000.162 ±\pm 0.0420.261 ±\pm 0.113
ResNet180.338 ±\pm 0.0410.324 ±\pm 0.0140.295 ±\pm 0.0600.333 ±\pm 0.0660.231 ±\pm 0.0400.093 ±\pm 0.0620.109 ±\pm 0.0350.380 ±\pm 0.0020.153 ±\pm 0.0550.251 ±\pm 0.108
ResNet340.347 ±\pm 0.0440.327 ±\pm 0.0170.307 ±\pm 0.0720.315 ±\pm 0.0590.219 ±\pm 0.0310.068 ±\pm 0.0460.115 ±\pm 0.0330.371 ±\pm 0.0230.154 ±\pm 0.0470.247 ±\pm 0.111
ResNet500.341 ±\pm 0.0320.331 ±\pm 0.0150.319 ±\pm 0.0670.361 ±\pm 0.1170.212 ±\pm 0.0130.097 ±\pm 0.0510.116 ±\pm 0.0320.359 ±\pm 0.0300.157 ±\pm 0.0440.255 ±\pm 0.109
ViT-b/160.326 ±\pm 0.0310.339 ±\pm 0.0170.278 ±\pm 0.0430.302 ±\pm 0.0630.198 ±\pm 0.0020.110 ±\pm 0.0310.107 ±\pm 0.0260.363 ±\pm 0.0070.149 ±\pm 0.0480.241 ±\pm 0.101
Appendix
Table 9: Pearson Correlation Coefficient (PCC) across datasets and timm encoders for HEST-1k and HEST-1k-1NN . The best and the second best score per dataset are in bold and underlined respectively. Standard deviations across the cross-validation folds are reported.
HyperparameterValueMean Recall@5%
aggregationavg0.19 ±\pm 0.02
 cls0.20 ±\pm 0.02
sizes0.20 ±\pm 0.02
 b0.20 ±\pm 0.02
 l0.19 ±\pm 0.02
 h0.19 ±\pm 0.02
Appendix
Table 10: Hyperparameter sweep for untrained ViTs on the RxRx3-core benchmark. Values report mean Recall@5% ±\pm standard deviation across three folds.
ModelArchitectureFoldCORUMHuMAPReactomeSIGNORStringDB
SingleConvCNN10.284 ±\pm 0.0090.363 ±\pm 0.0070.080 ±\pm 0.0010.037 ±\pm 0.0080.222 ±\pm 0.003
SingleConvCNN20.281 ±\pm 0.0210.350 ±\pm 0.0300.064 ±\pm 0.0010.066 ±\pm 0.0050.226 ±\pm 0.011
SingleConvCNN30.263 ±\pm 0.0040.309 ±\pm 0.0140.068 ±\pm 0.0010.046 ±\pm 0.0050.204 ±\pm 0.005
AlexNetCNN10.111 ±\pm 0.0390.106 ±\pm 0.0290.059 ±\pm 0.0050.057 ±\pm 0.0030.087 ±\pm 0.016
AlexNetCNN20.114 ±\pm 0.0120.120 ±\pm 0.0060.059 ±\pm 0.0170.057 ±\pm 0.0130.102 ±\pm 0.015
AlexNetCNN30.103 ±\pm 0.0230.098 ±\pm 0.0230.049 ±\pm 0.0100.058 ±\pm 0.0180.089 ±\pm 0.021
Appendix
Table 11: Impact of weight initialization on recall at 5% across biological benchmarks. Mean and standard deviation across 3 random seeds for untrained models. Results are reported for all benchmarks (CORUM, HuMAP, Reactome, SIGNOR, StringDB) across three evaluation folds.
ModelArchitectureFoldMean mAP
SingleConvCNN00.357 ±\pm0.018
SingleConvCNN10.315 ±\pm0.008
SingleConvCNN20.366 ±\pm0.017
SingleConvCNN30.340 ±\pm0.016
SingleConvCNN40.341 ±\pm0.023
ResNet18CNN00.146 ±\pm0.002
Appendix
Table 12: Per-fold compound retrieval performance on the JUMP-CP subset benchmark . Mean mAP is reported as mean ±\pm standard deviation across 3 random seeds.
Feature typeModelAverage Recall@5%
Pixels statisticsPixel Stats0.16±0.020.16\pm 0.02
Human handcrafted featuresCellProfiler0.17±0.010.17\pm 0.01
Untrained networkViT-l/160.19±0.020.19\pm 0.02
Pretrained Fondation ModelOpenPhenom0.20±0.020.20\pm 0.02
Appendix
Table 13: RxRx3-core retrieval performance of CellProfiler features compared with pixel statistics, the best untrained ViT model, and OpenPhenom. Values report mean Recall@5% ±\pm standard deviation across three folds.
Figure 18: Recall @5% for all models on the original full RxRx3-Core dataset.
Figure 19: Model performance as a function of network stage. Both panels evaluate embeddings at four intermediate stages including the last layer. (a) Evolution of micro-accuracy on the ImageNet-1k validation set. (b) Average recall across all RxRx3-Core evaluated datasets.
Figure 20: Representational dimensionality versus performance on biological and natural image benchmarks . Dataset specific scores are plotted against variance explained by the first two PCA components. Right axis: HuMAP mean recall at top 5%. Left axis: ImageNet-1k kNN top-1 accuracy. Points correspond to different model architectures, training states, and extraction layers.
Figure 21: Impact of selected pairs % on the recall per dataset. A few important and diverse models are displayed. Red dash line indicates the random selection
Figure 22: Benchmarking model performance across cell culture and natural images tasks. Performance is evaluated across: (a) gene-gene interaction retrieval and (b) ImageNet-1k classification. Solid bars indicate pretrained models; hatched bars indicate untrained models. Error bars represent variability across evaluation folds, and dashed horizontal lines denote random baselines.
Figure 23: Model performance on gene-gene interaction retrieval when looking at bottom 5%5\% of cosine similarities between gene-gene pairs.
Figure 24: Representational Similarity Analysis (RSA) across model architectures. The heatmap displays the pairwise similarity between model embeddings calculated using Spearman’s rank correlation coefficient. Models are hierarchically clustered to reveal functional groupings. Labels prefixed with [P] (blue) denote pretrained models, while [U] (red) indicates untrained/random initializations.
Figure 25: PCA Variance Explained of RxRx3-Core vs. Recall @5%@5\% across various datasets (CORUM, StringDB, Reactome and SIGNOR).
Figure 26: Evolution of representation quality across intermediate layers for JUMP-CP compounds. Performance is evaluated as Mean Average Precision (mAP) for ImageNet retrieval across eight distinct positive control compound. Each panel displays the mAP score as a function of the network intermediate layers (Layer Depth). Solid lines denote pretrained models, while dashed lines indicate untrained models; shaded areas represent variability across a given architecture. The dashed horizontal line in each plot represents the theoretical random baseline ( 0.1250.125 ).
ModelMean % of ReplicatingMean % of Matching
Random5.05.0
Pixel Stats22.310.5
OpenPhenom37.811.4
ResNet152 (untrained)27.311.6
ViT-l/16 (pretrained)36.812.3
SingleConv (untrained)30.512.8
Appendix
Table 14: JUMP-MoA retrieval results. Means across 86 compounds for % of replicating and across 43 MoAs for % of matching.
  DatasetBACHBRACS
ModalityTrainingModel  
Morphological cell graphsPretrained IIDACM-bio38.37 ±\pm 2.1621.55 ±\pm 2.62
ACM-UNI29.03 ±\pm 9.9219.95 ±\pm 3.27
  GrapHist69.16 ±\pm 3.3760.30 ±\pm 0.46
Full RGB ImagesPretrained, IIDDINOv259.25 ±\pm 3.3452.68 ±\pm 1.78
MAE57.94 ±\pm 3.9156.33 ±\pm 0.76
Appendix
Table 16: Reported reference results from ( Öğüt et al., 2026 ) . The RoI-level datasets are processed in patches and aggregated using three different MIL approaches. The provided values correspond to the best MIL strategy per model. Held-out test macro F1 scores are reported with standard deviations across models fit on 5 cross-validation folds.
  DatasetBACHBRACS (448 ×\times 448)BRACS (896 ×\times 896)
ModalityModelTraining   
Full RGB ImagesH-Optimus-1Pretrained, IID90.88 ±\pm 2.1348.20 ±\pm 1.7175.65 ±\pm 0.99
Morphological Featuresmean_std-62.06 ±\pm 2.4256.98 ±\pm 1.0258.02 ±\pm 1.17
Structure Only (Images of Binary Graphs)ResNet152Pretrained, OOD54.83 ±\pm 2.8647.29 ±\pm 1.9853.32 ±\pm 1.37
ResNet18Pretrained, OOD48.21 ±\pm 4.2243.31 ±\pm 1.5646.90 ±\pm 1.71
ResNet34Pretrained, OOD53.43 ±\pm 2.2244.95 ±\pm 0.7045.72 ±\pm 1.83
Appendix
Table 17: The best and the second best score per dataset are in bold and underlined respectively. Following the reference implementation ( Öğüt et al., 2026 ) we provide the macro F1 scores and standard deviations across five validation folds in parentheses. For all datasets the best performing patch aggregation method is reported. For BACH the best performing patchification strategy is reported.
More in this topic

Representation Learning

Representation learning develops data transformations that expose useful structure for prediction, transfer, generation, and control.

See all 804 papers

Explore similar work

CardsList

  1. MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays

    May 14, 2026Emre Hayir, Lorin Crawford, Alex X. LuMicroscopy Image Analysis

    Microsoft Research New England

  2. Evaluating Vision Foundation Models for Pixel and Object Classification in Microscopy

    Mar 20, 2026Carolin Teuber, Anwai Archit, Tobias Boothe +3Vision Foundation ModelsSemantic Segmentation

    Georg-August-University Göttingen, Institute of Computer Science · Department of Tissue Dynamics and Regeneration, Max Planck Institute for Multidisciplinary Sciences, Göttingen · Georg-August-University Göttingen, Faculty of Biology and Psychology +3

  3. scMIR: a vision-language foundation model for single-cell light microscopy image representation

    Jul 21, 2026Yifan Shang, Jiahui Tan, Xiangxiang Zeng +1Cross-Modal Representation LearningVisual Representation Learning

    Laser Metrology and Biomedicine Lab, Department of Biomedical Engineering, The Chinese University of Hong Kong, Hong Kong, China · State Key Laboratory of Chemo and Biosensing, College of Computer Science and Electronic Engineering, Hunan University, Changsha 410082, China