ImmuVis: Hyperconvolutional Foundation Model for Imaging Mass Cytometry
Authors: Dawid Uchal, Marcin Możejko, Krzysztof Gogolewski, Piotr Kupidura, Fabian Ozga, Szymon Łukasik, Jakub Giezgała, Tomasz Nocoń, Kacper Pietrzyk, Robert Pieniuta, Mateusz Sulimowicz, Michał Zmysłowski, Michał Orzyłowski, Tomasz Siłkowski, Karol Zagródka, Eike Staub, Ewa Szczurek
Organizations: Faculty of Mathematics, Informatics and Mechanics, University of Warsaw, Warsaw, Poland · Merck Healthcare KGaA, Darmstadt, Germany · Institute of AI for Health, Helmholtz Munich, Neuherberg, Germany
We present ImmuVis, a family of efficient foundation models for imaging mass cytometry (IMC), a high-throughput multiplex imaging technology that handles molecular marker measurements as image channels and enables large-scale spatial tissue profiling. Unlike natural images, multiplex imaging lacks a fixed channel space, as real-world marker sets vary across studies, violating a core assumption of standard vision backbones. To address this, ImmuVis introduces marker-adaptive hyperconvolutions that generate convolutional kernels from learned marker embeddings, enabling a single model to operate on arbitrary measured marker subsets without retraining. We pretrain ImmuVis on the largest dataset to date, IMC17M (28 cohorts, 24,405 images, 265 markers, over 17M patches), using self-supervised masked reconstruction. ImmuVis outperforms state-of-the-art baselines and ablations in virtual staining and downstream classification tasks at substantially lower compute cost than transformer-based alternatives, and is the sole model that provides calibrated uncertainty via a heteroscedastic likelihood objective. These results position ImmuVis as a practical framework for real-world IMC modeling.
Figures & tables
Figure 1: Motivation for ImmuVis : real-world panel heterogeneity. IMC17M exhibits strong cohort-marker diversity (rings: cohorts; radii: markers; colored ticks: measured markers), motivating a single model that operates on different marker subsets. ImmuVis encodes an observed panel (e.g., MITF, p53, pNFkB) into a shared pan-marker latent space for downstream phenotyping (cell typing, clinical prediction) and instantiates a task-specific decoder to virtually stain requested targets (e.g., CD45RO), outputting both expression and predictive uncertainty.
Figure 2: ImmuVis architecture overview. Marker-agnostic encoder stems embed each input marker channel and a hyperconvolution module, conditioned on learned marker embeddings, fuses them into a shared pan-marker representation processed by a standard vision backbone. A symmetric hyperconvolution and marker-agnostic decoder instantiates the requested output marker set, predicting per-marker reconstructions together with pixel-wise uncertainty (modified heteroscedastic log-variance).
Panel
EVA
ImmuVisViT
ImmuVisConv
ImmuVisSwin
VirTues
Head&Neck, Immucan [ 38 ]
.180/.51
.081/.72
.068/.77
.061 / .78
.101/.73
Allam et al. [ 39 ] , Panel 1
.067/.50
.045/.64
.037/ .66
.044/ .66
.035 /.65
Allam et al. [ 39 ] , Panel 2
.462/.47
.287/.56
.217/.62
.187/.59
.185 / .65
Hu et al. [ 40 ] , Panel 1
.108/.18
.113/.32
.092/.41
.081/ .45
.056 /.34
Hu et al. [ 40 ] , Panel 2
.487/.47
.457/.48
.427 / .61
.531/.45
.455/.56
Hu et al. [ 40 ] , Panel 3
.021/.54
.010/.58
.017/.53
.005 /.55
.011/ .65
Table 1: Virtual staining on ten test panels: MSE/Pearson. VirTues and EVA denote their IMC17M+OP variants with overlap-only inputs (full-panel comparison: Table 11 ). Bold marks the lowest MSE and highest Pearson per panel, including ties at the reported precision.
Model
Pearson ( ↑ )
ECE ( ↓ )
AUSE ( ↓ )
ImmuVisViT
.922
.029
.0135
ImmuVisConv
.941
.029
.0132
ImmuVisSwin
.934
.029
.0112
Table 2: Uncertainty calibration evaluation on the HN cohort , measured by Pearson correlation ( predicted uncertainty vs reconstruction error), expected calibration error (ECE), and area under sparsification error (AUSE, RMSE-base). Bold denotes the best score; ECE values are equal at the reported precision.
Model
( 46 )
[ 47 ]
[ 48 ]
Avg.
mean-markerraw
.715
.535
.684
.645
mean-markerOP
.778
.563
.705
.682
mean-markerVP
.762
.565
.706
.678
EVAOP
.602
.510
.691
.601
EVAIMC17M+OP
.694
.551
.697
.647
VirTues
.726
.569*
.711
.669
Table 3: Cell-typing Macro-F1 from 10-fold cross-validation. All standard deviations were below 10−2 . For VirTues, values report the better score from either the original or ImmuVis cell-typing procedure (see Appendix A.17 ). Underline denotes the best score within a model family; bold denotes the best score per endpoint.
Model
Danenberg et al. [ 47 ]
Cords et al. [ 48 ]
PAM50
Grade
ER
ERBB2
Subtype
Relapse
Grade
Avg.
mean-markerraw
.41±.06
.47±.05
.77±.08
.75±.11
.77±.03
.56±.02
.49±.05
.602
mean-markerOP
.42±.06
.49±.06
.76±.09
.75±.11
.78±.03
.56±.02
.49±.04
.606
mean-markerVP
.43±.05
.47±.05
.75±.09
.77±.10
.79±.03
.55±.03
.49±.05
.608
EVAOP
.38±.05
.50±.08
.75±.07
.61±.08
.79±.03
.57±.05
.50±.05
.585
EVAIMC17M+OP
.34±.06
.43±.05
.68±.05
.47±.002
.79±.03
.58±.04
.51±.08
.543
Table 4: Representation learning for clinical endpoints (Macro-F1; higher is better). Values report mean ± std Macro-F1 across 10-fold cross-validation; bold marks the highest displayed mean, including rounded ties.
Model
Active param.
GFLOPs
Time [ms]
Mem. [GB]
ImmuVisConv
44.6M
89.9
13.4 ± 0.2
1.01
ImmuVisSwin
57.0M
138.8
18.2 ± 0.9
1.03
ImmuVisViT
105.2M
141.5
11.1 ± 0.3
1.39
VirTues
42.5M
1049.3
28.5 ± 0.5
0.45
EVA
122.4M
3590.0
45.8 ± 0.3
0.95
Table 5: Computational efficiency. Inference cost measured on 100 random inputs of shape [40,128,128] on a single NVIDIA RTX 5000 Ada Generation GPU.
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Subjects
Images
Markers
Histopathology
VirTues
ajaib [ 50 ]
5
34
34
Glioblastoma
bengsch [ 51 ]
54
108
42
Hepatocellular carcinoma
bowen-lynch [ 52 ]
18
85
38
Lynch Syndrome colorectal mucosa
cho [ 53 ]
12
140
41
Pancreatic adenocarcinoma
cords [ 48 ]
1070
2070
42
Non-small cell lung cancer
✓
cords-fibro [ 54 ]
12
110
42
Breast cancer
✓
Appendix
Table 6: Cohort-level composition of IMC17M . The dataset spans 28 cohorts, 24,405 images, 265 markers, and over 17M patches. IMMUcan BC2/NSCLC2 (P1/2)/SCCHN1 datasets are IMMUcan [ 38 ] proprietary data, awaiting publication. The VirTues column denotes with ✓cohorts included in the VirTues pretraining corpus.
Component
Parameter
Value / Specification
ImmuVisConv
Encoder (Marker-Agnostic)
Stages ( Lma )
1
Stage composition
4-fold Downsampling + 4 ConvNeXt v2 blocks
Stage dimension ( dma )
16
Encoder (Pan-Marker)
Hyperkernel size ( he×we )
1×1
Pan-Marker dimension ( dpm )
192
Appendix
Table 7: ImmuVis models architecture and parameter specifications
Lma
Validation NLL
Cell typing
Staining MSE
0 (fixed-channel control)
−6.601
.758
.0672
1 (used)
−7.410
.790
.0573
2
−7.546
.788
.0833
Appendix
Table 8: Conv stem ablation. Cell typing is Macro-F1; lower validation NLL and staining MSE are better.
Figure 3: MSE of VirTues reconstructions in trained vs zero-shot settings.
Figure 4: Examples of VirTues reconstructions for trained and zero-shot settings.
Figure 5: Exemplary qualitative comparison of spatial preprocessing on 32×32 IMC patches from [ 46 ] . Each row shows a single antibody channel: E-cadherin (Ecad, epithelial marker), CD3 (T-cell marker), and PD-1. The six columns are: raw counts; arcsinh (x/5) normalisation; the corresponding omnidirectional gradient m2 and absolute curvature ∣m1∣ maps (computed on the arcsinh-normalised full image and coloured with sequential blue and green scales, respectively); arcsinh + Gaussian blur ( σ=1 px, 3×3 kernel, analogous to VirTues preprocessing); and arcsinh + Butterworth low-pass filter (order 2, cutoff 0.2 Nyquist, used by ImmuVis ). Intensity columns share a common colour scale across arcsinh and filtered variants within each row; the raw column and the two metric columns are each scaled independently per row.
Figure 6: Joint distribution of curvature ∣m1∣ and gradient m2 under six preprocessing pipelines. Rows: no spatial filter / Butterworth / Gaussian; columns: arcsinh (x/5) / log(1+x) normalisation. Each panel accumulates a 2D histogram over all pixels, images, and antibody channels, using a mixed bin scale (one zero-anchor bin [0,10−4] followed by 99 log-spaced bins up to 10 ). All six panels share a logarithmic colour scale. The white curve shows E[∣m1∣∣m2] ; the diagonal ∣m1∣=m2 is for a reference. Both filters reduce pixel curvature relative to the unfiltered signal (top row); Gaussian smoothing introduces discrete curvature artefacts at fixed ∣m1∣ values, whereas Butterworth low-pass filtering yields a smoother, more uniform (∣m1∣,m2) distribution.
Setting
P1
P2
MSE
Pearson
No adaptation
.765
.643
–
–
New-marker embeddings only
.765
.641
.081
.531
Cohort included in pretraining
.776
.649
.040
.687
Original VirTues
.734
.617
–
–
Appendix
Table 10: Moldoveanu transfer with ImmuVisConv : cell-typing Macro-F1 for panels P1/P2 and staining on all 48 markers where available. The cohort was included in original VirTues pretraining.
Panel
VirTues F
VirTues O
EVA F
EVA O
A1
.0348/.62
.0350/.65
.0659/.50
.0672/.50
A2
.1853/.65
.1853/.65
.3498/.48
.4618/.47
H1
.0661/.29
.0559/.34
.1060/.18
.1080/.18
M1
.2634/.52
.2011/.54
.3577/.37
.3612/.39
M2
.2299/.24
.1849/.29
.3112/.20
.3019/.21
Appendix
Table 11: MSE/Pearson on shared targets with full-panel (F) and overlap-only (O) inputs.
Model
MSE ( ↓ )
Pearson ( ↑ )
VirTues
.797
.700
VirTuesIMC17M
.410
.764
ImmuVisViT
.502
.731
ImmuVisConv
.400
.783
Appendix
Table 12: Virtual staining accuracy measured by MSE and Pearson correlation under the VirTues preprocessing protocol. Bold denotes the best score.
Figure 7: Quantitative virtual staining evaluation for MSE on the Head & Neck cohort for models train using VirTues preprocessing ( VP ). Per-marker reconstruction accuracy measured by image-level log(MSE) for ImmuVisConv , ImmuVisViT , VirTuesIMC17M and VirTues (original checkpoint from [ 13 ] ). Each boxplot summarizes the distribution of image-level scores, where each image score is obtained by averaging patch-level errors over all patches from that image. The three rows above the plot report paired Wilcoxon signed-rank tests results across images for the corresponding model pairs (as indicated in the left top corner), with significance after FDR correction (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 8: Quantitative virtual staining evaluation for Pearson correlation on the Head & Neck cohort for models trained using VirTues preprocessing ( VP ). Per-model Pearson correlation measured on a marker-level for ImmuVisConv , ImmuVisViT , VirTuesIMC17M and VirTues (original checkpoint from [ 13 ] ). Each boxplot summarizes the distribution of marker-level Pearson correlation computed across all images. P-values are obtained via Wilcoxon signed-rank tests results across markers for the corresponding model pairs (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 9: Quantitative virtual staining evaluation for MSE on the Head & Neck cohort for models train using ImmuVis preprocessing ( OP ). Per-marker reconstruction accuracy measured by image-level log(MSE) for ImmuVisConv , ImmuVisSwin , ImmuVisViT , and VirTuesIMC17M+OP . Each boxplot summarizes the distribution of image-level scores, where each image score is obtained by averaging patch-level errors over all patches from that image. The three rows above the plot report paired Wilcoxon signed-rank tests results across images for the corresponding model pairs (as indicated in the left top corner), with significance after FDR correction (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 10: Quantitative virtual staining evaluation for Pearson correlation on the Head & Neck cohort for models train using ImmuVis preprocessing ( OP ). Per-model Pearson correlation measured on a marker-level for ImmuVisConv , ImmuVisSwin , ImmuVisViT , VirTuesIMC17M+OP and EVAOP (released checkpoint evaluated outside its native preprocessing). Each boxplot summarizes the distribution of marker-level Pearson correlation computed across all images. P-values are obtained via Wilcoxon signed-rank tests results across markers for the corresponding model pairs (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 11: Predicted uncertainty tracks reconstruction error across (image, channel) pairs under out-of-cohort LOO virtual staining. Each point is one LOO reconstruction (5 960 per model). Strong log–log scaling between mean predicted variance and MAE holds for all three ImmuVis variants: Pearson rCONV=.941 , rViT=.922 , rSwin=.934 . The OLS fit slopes ( ImmuVisConv , ImmuVisViT , ImmuVisSwin ) are .57 , .60 , .54 respectively. The relationship confirms that the heteroscedastic head captures local ambiguity in the reconstruction.
Figure 12: Reliability diagram - global, out-of-cohort LOO. Empirical coverage of the predictive Gaussian’s central α -interval as a function of nominal α . All three ImmuVis variants closely track the perfect-calibration line and their ECE values agree to within .0006 . The inset, plotted on a magnified y -axis, reveals that the three architectures miscalibrate in qualitatively different directions: ImmuVisConv predominantly under-covers (peaking near α≈0.3 ), while ImmuVisViT and ImmuVisSwin predominantly over-cover (peaking near α≈0.7 , with ImmuVisSwin deviating most). Integrating the absolute deviations gives nearly identical ECE values ( .029±.0003 ), indicating that overall calibration quality is comparable across architectures despite distinct miscalibration shapes.
Figure 13: Sharpness vs. ECE per marker. Each point is one of the 40 markers in the Head & Neck cohort. The bulk of markers cluster in the lower-left quadrant (mean σ≲.05 , ECE <.05 ), confirming that the calibrated coverage in Figure 12 is achieved at narrow predictive intervals.
Figure 14: Sparsification analysis - global, out-of-cohort LOO (RMSE base, matches Table 2 ). Left: residual RMSE on the kept (1−f) fraction of pixels after removing the most-uncertain f by predicted σ (coloured lines per model); the grey oracle curve is obtained by ranking with ∣r∣ instead and represents the lowest achievable curve at this residual distribution. AUSE in the legend is the area between each model’s curve and the oracle. Right: the sparsification-error integrand SE(f)=errσ(f)−erroracle(f) . All three ImmuVis variants track the oracle closely across the full range; the residual gap is concentrated in the highest- σ tail ( f≲0.2 ), where outlier pixels are hardest to rank exactly. ImmuVisSwin achieves the lowest AUSE, while ImmuVisViT exhibits the largest peak sparsification error.
Figure 15: Same construction as Figure 14 with MAE in place of RMSE. The model ranking is preserved ( ImmuVisSwin < ImmuVisConv < ImmuVisViT ), confirming the σ -ordering quality differences are robust to the error norm choice.
Model
[ 46 ]
[ 47 ]
[ 48 ]
Avg.
Method
patch
context
patch
context
patch
context
patch
context
EVAOP
.602
.450
.510
.487
.691
.571
.601
.503
VirTues
.726
.675
.555
.569
.711
.699
.664
.648
VirTuesIMC17M
.703
.672
.517
.555
.712
.704
.644
.644
VirTuesIMC17M+OP
.706
.686
.557
.579
.716
.702
.660
.656
ImmuVisViT
.753
.759
.592
.618
.751
.740
.699
.706
Appendix
Table 13: Representation learning performance for cell typing. Macro-F1 scores are reported for 10-fold cross-validation; higher values indicate better performance. Standard deviations across folds were below 10−2 for all entries and are therefore omitted. The best score for each endpoint is shown in bold . The best method per endpoint / method is underlined .
Model
[ 46 ]
[ 47 ]
[ 48 ]
Avg.
EVAOP
.664
.559
.748
.657
VirTues
.812
.635*
.781
.743
VirTuesIMC17M
.786
.624*
.783
.731
VirTuesIMC17M+OP
.796
.651*
.783
.743
ImmuVisViT
.845
.662
.827
.778
ImmuVisConv
.882
.668
.822
.791
Appendix
Table 14: Representation learning performance for cell typing (Average Precision; higher is better). Scores are reported from 10-fold cross-validation. Standard deviations across folds were below 10−2 for all entries and are therefore omitted. For VirTues evaluations, the reported value is the higher score obtained using either the authors’ original or the ImmuVis cell-typing approach (see A.17 ). ( best score within models family, best score for each endpoint)
Model
[ 46 ]
[ 47 ]
[ 48 ]
Avg.
Method
patch
context
patch
context
patch
context
patch
context
EVAOP
.664
.472
.559
.534
.748
.606
.657
.537
VirTues
.812
.760
.625
.635
.781
.776
.739
.724
VirTuesIMC17M
.786
.756
.574
.624
.783
.778
.714
.719
VirTuesIMC17M+OP
.796
.771
.624
.651
.783
.771
.734
.731
ImmuVisViT
.845
.848
.662
.695
.827
.821
.778
.788
Appendix
Table 15: Representation learning performance for cell typing. ImmuVisConv context entries use the original 8×8 protocol (see text for corrected Eling Macro-F1). Average Precision scores are reported for 10-fold cross-validation; higher values indicate better performance. Standard deviations across folds were below 10−2 for all entries and are therefore omitted. The best score for each endpoint is shown in bold . The best method per endpoint / method is underlined .
Figure 16: Cell-typing performance from learned representations on the [ 46 ] dataset. Bars show mean marco-F1 across cross-validation folds with confidence intervals (whiskers); higher is better. The seven rows above the plot report paired Wilcoxon signed-rank tests on per-fold scores for the corresponding model pairs (indicated by the colored squares on the left). Significance markers are color-coded to match the model with the higher mean score. Significance is shown after FDR correction (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 17: Cell-typing performance from learned representations on the [ 47 ] dataset. Bars show mean marco-F1 across cross-validation folds with confidence intervals (whiskers); higher is better. The seven rows above the plot report paired Wilcoxon signed-rank tests on per-fold scores for the corresponding model pairs (indicated by the colored squares on the left). Significance markers are color-coded to match the model with the higher mean score. Significance is shown after FDR correction (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Figure 18: Cell-typing performance from learned representations on the [ 48 ] dataset. Bars show mean marco-F1 across cross-validation folds with confidence intervals (whiskers); higher is better. The seven rows above the plot report paired Wilcoxon signed-rank tests on per-fold scores for the corresponding model pairs (indicated by the colored squares on the left). Significance markers are color-coded to match the model with the higher mean score. Significance is shown after FDR correction (ns - not significant; (∗)<0.05 ; (∗∗)<0.01 ; (∗∗∗)<0.001 ).
Model
Probe
Danenberg et al. [ 47 ]
Cords et al. [ 48 ]
PAM50
Grade
ER
ERBB2
Subtype
Relapse
Grade
Avg.
EVAOP
LogReg
.38±.05
.50±.08
.75±.07
.61±.08
.79±.03
.57±.05
.50±.05
.585
abMIL
.40±.03
.45±.03
.80±.07
.81±.15
.81±.02
.55±.05
.48±.04
.614
VirTues
LogReg
.37±.05
.53±.05
.78±.08
.76±.11
.82±.03
.57±.03
.51±.07
.620
abMIL
.42±.06
.48±.04
.77±.09
.66±.11
.83±.02
.58±0.03
.48±.03
.603
VirTuesIMC17M
LogReg
.36±.06
.49±.05
.77±.06
.78±.13
.83±.02
.58±.04
.54±.06
.621
Appendix
Table 16: Representation learning for clinical endpoints (Macro-F1; higher is better). Values report mean ± std Macro-F1 across 10-fold cross-validation; best performance per endpoint is shown in bold.
Figure 19: Locality index analysis on HN. Distributions of the mean L∞ saliency radius for ImmuVis variants and VirTues. Smaller values indicate more spatially local influence. Pairwise comparisons use paired Wilcoxon signed-rank tests (***: p<0.001 ).
Model
Immucan H&N
Cords
Danenberg
ImmuVisConv
26.0
26.1
24.1
ImmuVisSwin
28.5
26.6
25.8
ImmuVisViT
29.9
29.8
29.1
VirTues
36.5
37.9
37.8
Appendix
Table 17: Mean saliency locality radius on three cohorts; smaller values indicate more local influence.
Cytometry measures the complex characteristics of single cells (e.g., counts and protein expression of immune cells) and is widely used across immunological research and clinical settings. However, cytometry data is highly heterogeneous and unstandardized due to experimental protocols and the choice of measured features. While machine learning methods hold the potential to gain deeper insights into cell biology, these challenges make them difficult to apply and transfer across studies. Recent advances in foundation models can alleviate these issues, but corresponding approaches are still scarce in this field. To address this, we provide CytoBERT, a publicly available, open-source, open-weight foundation model for single-cell cytometry data with variable marker panels. CytoBERT is pretrained in a self-supervised manner on a large-scale cytometry corpus (15 human datasets with heterogeneous marker panels and more than 50 million cells) curated through marker standardization, enabling it to learn transferable inter-marker relationships within cells. Fine-tuning CytoBERT for sample-level classification demonstrates that transfer learning across heterogeneous cytometry datasets is feasible, providing a starting point for scalable, generalizable cytometry analysis. Code is available at GitHub.
Syed Abdul Haseeb Qadri, Bjarne C. Hiller, Felix Blanke +7
University of Rostock, Rostock, Germany · Marburg University, Marburg, Germany · Hessian Center for Artificial Intelligence, Darmstadt, Germany
Understanding tissue organisation in multiplexed imaging requires modelling both cellular phenotypes and their spatial context. Existing approaches typically rely on handcrafted features, such as marker intensity statistics or cell-type proportions, which often fail to scale or generalise across cohorts with heterogeneous marker panels. We introduce MUL-T, a lightweight transformer framework that reframes tissue architecture as a masked contextual prediction task over discrete cell tokens. By learning contextualised [CLS] embeddings without task-specific supervision, the model captures higher-order cellular interactions while remaining computationally efficient. We evaluate MUL-T on several clinically relevant downstream tasks, including core-level tumour pattern classification, patient-level grading, PD-L1 positivity prediction, and cross-dataset treatment response prediction. Across tasks, MUL-T consistently outperforms classical feature-based baselines and achieves performance comparable to a foundation ViT model, despite substantially fewer parameters and lower training cost.
Farzaneh Seyedshahi, Kai Rakovic, Adalberto Claudio Quiros +2
Cancer Research UK Scotland Institute, Glasgow, UK · University of Glasgow, Glasgow, UK
Multiplex immunofluorescence (mIF) enables simultaneous single-cell quantification of multiple biomarkers within intact tissue architecture, yet its high reagent cost, multi-round staining protocols, and need for specialized imaging platforms limit routine clinical adoption. Virtual staining can synthesize mIF channels from widely available brightfield immunohistochemistry (IHC), but current translators optimize pixel-level fidelity without explicitly constraining nuclear morphology. In pathology, this gap is clinically consequential: subtle distortions in nuclei count, shape, or spatial arrangement propagate directly to quantification endpoints such as the Ki67 proliferation index, where errors of a few percent can shift treatment-relevant risk categories. This work introduces a supervision-free, architecture-agnostic conditioning strategy that injects a continuous cell probability map from a pretrained nuclei segmentation foundation model as an explicit input prior, together with a variance-preserving regularization term that matches local intensity statistics to maintain cell-level heterogeneity in synthesized fluorescence channels. The soft prior retains gradient-level boundary information lost by binary thresholding, providing a richer conditioning signal without task-specific tuning. Controlled experiments across Pix2Pix with U-Net and ResNet generators, deterministic regression U-Net, and conditional diffusion on two independent datasets demonstrate consistent improvements in nuclei count fidelity and perceptual quality, as the sole modifications. Code will be made publicly available upon acceptance.
Junhyeok Lee, Han Jang, Heeseong Eum +2
Interdisciplinary Program in Cancer Biology, Seoul National University College of Medicine, Seoul, Republic of Korea · Interdisciplinary Program in Bioengineering, Seoul National University, Seoul, Republic of Korea · Department of Biomedical Sciences, Seoul National University, Seoul, Republic of Korea +3