List of Papers Browse by Subject Areas Author List
Abstract
Pervasive batch effects are a common issue, especially in recent large-scale Cell Painting datasets, which have been produced to aid AI-enhanced drug discovery efforts. Technical differences arising from experiments carried out in different batches can cause models to fail to generalize to unseen batches, despite good predictive performance “within batch”. We propose a biologically grounded test-time adaptation framework, SHOT-CCR, which uses cell-invariant gradient reversal to decouple morphological signal from experimental confounders. Our approach performs 4.5% better than the current RxRx1 benchmark, classifying 1,139 classes of siRNA genetic perturbations with 91.6% accuracy. We deliver consistent results over four distinct cell types and two prominent Cell Painting datasets – RxRx1 and a subset of JUMP-CP. Across 484 classes of CRISPR perturbations in JUMP-CP our method improves accuracy by 15.7%. All code will be made publicly available upon publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4889_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/williamdee1/shot-ccr
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DeeWil_SHOTCCR_MICCAI2026,
author = { Dee, William AND Wenteler, Aaron AND Seal, Srijit AND Morris, Otto AND Slabaugh, Gregory},
title = { { SHOT-CCR: Biologically Guided Adversarial Training for Test-Time Adaptation in Cellular Morphology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces SHOT-CCR (SHOT Cell Count Reversal), a biologically informed test-time adaptation (TTA) framework for mitigating batch effects in Cell Painting microscopy data, with the goal of improving generalization of perturbation classification models to unseen experimental batches. The central technical innovation is a cell-count gradient reversal (CCR) mechanism integrated into the training pipeline. Rather than attempting to remove all batch-related information from learned representations, as generic domain-adversarial approaches do, SHOT-CCR selectively suppresses the influence of cell count, a biologically motivated confounder that varies systematically across experimental batches and cell types. This is achieved by attaching a cell count regression head to the feature extractor, with a gradient reversal layer that feeds back the inverse of the regression loss, discouraging the encoder from encoding cell-count-driven batch differences while preserving other batch structure that may still be predictive of the underlying genetic perturbation. This CCR mechanism is combined at test time with SHOT (Source Hypothesis Transfer), an unsupervised domain adaptation method that freezes the perturbation classifier and adapts only the feature extractor using an information maximization objective comprising entropy minimization, a diversity term, and pseudo-label classification loss. The architecture is built on a DenseNet-161 backbone and is evaluated across two large-scale Cell Painting datasets, namely RxRx1 and a curated subset of JUMP-CP, covering four cell types and two perturbation modalities (siRNA knockdown and CRISPR knockout). On RxRx1, the proposed method achieves 91.6% perturbation classification accuracy across 1,139 siRNA classes, a 4.5% improvement over the prior AdaBN benchmark. On the JUMP-CP subset (484 CRISPR perturbations), the method improves accuracy by 15.7% over AdaBN. The authors additionally release the curated JUMP-CP subset as a new public benchmark for the field.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Biologically motivated adversarial training. The key novelty, which is using cell count as a specific, biologically grounded adversarial target rather than a generic batch label, is well-motivated and conceptually elegant. Cell count is known to vary across batches and cell types and has been shown in prior work to be a strong predictor of bioactivity signals. By targeting this specific confounder rather than indiscriminately suppressing batch information, the method avoids the well-known pitfall of over-correction, which can strip task-relevant signal along with batch noise. The authors explicitly acknowledge and theoretically justify the use of partial invariance (citing Ben-David et al. and Ganin et al. ), which reflects genuine engagement with the domain adaptation literature. The decision to retain some batch structure while suppressing cell-count-driven variance is a principled design choice, not an ad hoc one. 2.Strong and well-contextualized empirical performance. The 4.5% improvement over the prior RxRx1 benchmark is statistically significant (t=63.64, p<0.0001), and the additional 1.6% gain from CCR over vanilla SHOT is also significant (t=9.74, p<0.0001). Gains are consistent across all four cell types, with the most pronounced improvement in U2OS (+8.0%), the most data-scarce and historically most challenging cell type. This is a meaningful result: it suggests the approach is particularly valuable precisely where data is hardest to obtain, which has practical relevance for drug discovery applications. Results are reported with variance over five seeds on RxRx1 and five cross-batch folds on JUMP-CP, lending credibility to the estimates. 3.Honest and insightful ablation analysis. The paper carefully decomposes contributions from Tent, SHOT, batch gradient reversal, and CCR individually (Table 2). Crucially, the authors do not overstate the contribution of CCR on JUMP-CP, where cell count distributions are homogeneous across batches and CCR provides only marginal benefit. This negative result is correctly interpreted as consistent with the hypothesis: CCR’s utility scales with cell count heterogeneity across batches. This kind of scientifically honest reporting, explaining why a method works or doesn’t in a given setting, strengthens the overall argument considerably. 4.Practical applicability and source-free adaptation. Unlike some competing approaches (e. g. , the generative domain-adversarial method of Yan et al. , 2025), SHOT-CCR requires no access to source data during test-time adaptation and no model retraining. This makes it directly applicable in realistic deployment scenarios where a pretrained model must generalize to new experimental batches without access to the original training data, which is a common situation in pharmaceutical and academic HCS pipelines. 5.Dual-dataset evaluation with distinct perturbation types. Evaluating on both RxRx1 (siRNA, multi-cell-type) and JUMP-CP (CRISPR, single cell type) provides meaningful breadth. The two datasets differ in perturbation modality, channel count, batch structure, and data volume, making the consistent (if context-dependent) gains more informative than a single-dataset result would be. The introduction of the curated JUMP-CP subset as a new benchmark is also a useful contribution to the community.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Limited novelty of individual components. Each of the three main building blocks, namely, AdaBN, Tent, and SHOT, is directly imported from prior computer vision literature with minimal modification. Gradient reversal itself is a well-established technique (Ganin et al. , 2017). The novelty lies in the combination and the choice of adversarial target (cell count rather than batch label), which is genuinely interesting, but the paper would benefit from a more explicit articulation of what is architecturally new versus what is application-domain transfer. Reviewers familiar with TTA and domain adversarial literature may view the contribution as primarily empirical rather than methodological. 2.Cell count as the sole biological confounder. The paper focuses exclusively on cell count as the adversarial target, justified by its established importance in the Cell Painting literature. However, it is well-known that other confounders, such as cell viability, plate position effects, fluorescence intensity drift, and reagent batch variability, also contribute substantially to batch effects in HCS data (Arevalo et al. , Nature Communications 2024; Ljosa et al. , Journal of Biomolecular Screening 2013). The generalizability of the CCR idea to other biological priors is noted in the conclusion but not demonstrated. It remains unclear whether cell count is uniquely effective or whether the gradient reversal framework would work equally well or better with other confounder signals. A brief ablation comparing CCR against reversal targeting another biologically motivated variable (e. g. , cell viability or plate position) would have substantially strengthened the claims. 3.JUMP-CP evaluation is limited in scope. The JUMP-CP subset used contains only five experimental batches and a single cell type (U2OS), and the authors acknowledge that the dataset is small and less diverse than RxRx1.The five-fold cross-batch evaluation is appropriate given the constraints, but the high variance in JUMP-CP results (e. g. , baseline: 10.4% ± 8.3%) makes it difficult to draw reliable conclusions about method behavior on this dataset. Furthermore, restricting JUMP-CP to genes also present in RxRx1 introduces a selection bias that may limit generalizability claims. Evaluation on a broader JUMP-CP subset, or on a third independent dataset, would have been more convincing. 4.Comparison with more recent baselines. The primary comparison is against the AdaBN method from Sypetkowski et al. (2023). While this is the established benchmark for RxRx1 and appropriate as a reference point, the field has moved forward. The recently published triple-effect correction method of Yan et al. (Nature Communications, 2025) and the masked autoencoder approach of Kraus et al. (CVPR 2024) represent potentially stronger baselines. The authors cite both but do not compare directly, noting that Yan et al. requires source data access. While this distinction is valid and the practical constraint is real, a direct numerical comparison, even under the note that operating conditions differ, would help situate SHOT-CCR more precisely within the current landscape. 5.Hyperparameter sensitivity is underexplored. The paper shows in Figure 3c that the cell count regression MSE loss varies substantially with the gradient reversal α parameter, and notes that separate learning rate and α values were “critical for model performance. “ However, the sensitivity analysis is limited to a few α values shown qualitatively. No systematic search or sensitivity table is provided for the full set of hyperparameters (α, regression learning rate, TTA learning rate, confidence threshold τ, early stopping patience). Given that the performance gains of CCR over vanilla SHOT are moderate (1.6% on RxRx1), it is important to establish that these gains are robust to reasonable hyperparameter variation and not artifacts of a well-tuned configuration on the test set. 6.No embedding-space analysis. The paper makes the claim that CCR selectively removes cell-count-driven batch differences while preserving other batch structure. This claim is asserted but not directly visualized. A UMAP or t-SNE analysis of feature embeddings before and after CCR training, colored by cell count, batch, or cell type, would provide direct evidence for this selective decorrelation and make the mechanistic story considerably more convincing. Such analyses are standard in the domain adaptation literature and their absence is notable.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a critical challenge in high-content screening: technical batch effects that hinder model generalization. The proposed SHOT-CCR framework introduces a novel cell-count gradient reversal (CCR) mechanism, which is a biologically grounded approach to domain adaptation. Rather than treating “batch” as a black box, the authors target a known confounder, namely, cell count, to preserve task-relevant morphological signals. This approach yields impressive results, setting a new benchmark on the RxRx1 dataset with 91.6% accuracy. Crucially, the method provides the largest gains (+8.0%) on the data-scarce U2OS cell type, suggesting it is highly effective where data is most difficult to obtain. Furthermore, its source-free nature makes it a practical solution for real-world pipelines where original training data may be inaccessible. However, the “Weak” qualifier is necessary because the core technical components, AdaBN, Tent, and SHOT, are directly imported from computer vision literature with minimal architectural modification. Additionally, the focus on a single confounder (cell count) and the absence of embedding-space visualizations (like UMAP) to confirm the selective decorrelation claim leave some mechanistic questions unanswered.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors propose a biologically grounded test-test adaption framework to decouple morphological signal from experimental confounders. They are provide a novel biologically guided adversarial mechanism that reduces batch effects.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- well written and clear introduction
- great practice to use the same training and test splits as the original paper for comparison
- table 2 - good experiments where we can compare all the different combinations of additions to the methodology
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- figure 2 - the pertained backbone lock icon is confusing - is it “unlocked” for inference as well? the weights are initialized by the pretrained weights but usually then frozen or “locked” during inference. In addition “At test time, the classifier is frozen” but it has a little lock sign also in the training
- no mention of publishing code
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
the manuscript is well written and organized. using established backbones and datasets makes the work comparable. With some clarifications on some of the comments above.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes SHOT-CCR, a test-time adaptation framework for Cell Painting data that combines SHOT-style source-free adaptation with adversarial training against a biologically motivated confounder, namely cell count. The idea is to reduce reliance on batch- and cell-type-correlated count information while preserving perturbation-relevant morphology. The method is evaluated on RxRx1 and a curated JUMP-CP subset and reports improved perturbation-classification performance over prior batch-correction baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper tackles a practically important problem in image-based profiling: batch effects that hinder generalization to new experimental batches.
The method is simple, intuitive, and practically useful. Using cell count as an explicit confounder is biologically motivated and more interpretable than generic domain-adversarial correction.
The empirical gains on RxRx1 are solid, and the paper also reports consistent improvement across all four cell types, including the more difficult U2OS setting.
The paper is generally well organized, and the experimental section is easy to follow. The ablation across Tent, SHOT, batch reversal, and CCR helps isolate where the gains come from.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper would benefit from stronger evidence that the gains truly come from removing a confounding signal rather than simply adding an auxiliary regularizer. For example, more analysis of representation quality, count-prediction leakage, or robustness to count-estimation noise would strengthen the claim.
The evaluation focuses primarily on perturbation classification accuracy. This is a reasonable and useful benchmark for batch generalization, but additional profiling-oriented evaluations, such as mechanism-of-action retrieval, compound similarity preservation, or downstream bioactivity prediction, would better demonstrate the broader utility of the learned representations for Cell Painting-based drug discovery.
The JUMP-CP experiment uses a custom filtered subset, which makes the practical value clear but reduces comparability with other methods.
The reliance on CellposeSAM-derived counts is reasonable, but the paper does not analyze how segmentation/counting errors affect adaptation.
Overall, the paper is strong empirically. Even though it is not primarily framed as a brand-new learning paradigm, it demonstrates that thoughtfully adapting modern TTA ideas to biomedical imaging can produce meaningful practical gains.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My overall score reflects a positive view of the paper’s empirical usefulness and clarity. I think the paper makes a meaningful practical contribution to Cell Painting batch adaptation, and the cell-count confounder is a sensible and well-motivated design choice. While the work builds on existing TTA ideas, I view that positively here: the authors translate and tailor them effectively to a biomedical imaging setting where robustness across batches and cell types is genuinely important. I am slightly above threshold because the problem is important, the results are convincing, and the method is likely to be useful in practice.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for their thoughtful and largely positive evaluations. We are encouraged that the reviewers found the work conceptually elegant and the results on RxRx1 and JUMP-CP significant. As an early accepted paper, we focus here on clarifying key points regarding novelty and methodology. Novelty of the CCR Mechanism (R1, R3) Reviewers noted that AdaBN, Tent, and SHOT are established methods. We agree and do not claim novelty in these components individually. The novel contribution is the cell-count gradient reversal (CCR) mechanism: the use of a biologically motivated, continuous confounder as the adversarial target rather than a discrete batch label. We found that generic batch-label adversarial training risks over-correction by suppressing perturbation-relevant variance which correlates with batch. CCR avoids this by selectively targeting the cell-count axis, which is a biologically relevant confounding variable that can dominate representations without uniquely identifying perturbation state. We will sharpen this framing in the camera-ready version. More broadly, our findings suggest that selective biologically guided invariance may be preferable to indiscriminate domain invariance in Cell Painting representation learning. Selective Decorrelation and Mechanistic Evidence (R1, R3) R1 and R3 request evidence to confirm selective decorrelation. We point to Table 2 as ablative evidence: batch gradient reversal, which targets a discrete batch label, produces a weaker result than CCR. This suggests cell-count specificity is critical and not a regularization artifact. Furthermore, JUMP-CP provides a natural negative control: where cell-count distributions are homogeneous across batches, CCR provides only marginal benefit, consistent with our hypothesis that its utility scales with confounder heterogeneity. Confounders and Clinical Relevance (R1, R3) While other confounders (e.g., plate position) exist, we focused on cell count as a fundamental biological prior proven to predict bioactivity. Our framework is extensible; the CCR head can be replaced or augmented with other regression targets. Crucially, the 8.0% gain on the U2OS cell type, the most data-scarce and challenging setting, demonstrates that our approach is most impactful where training data is hardest to obtain. Comparison with Recent Baselines (R1) Comparisons to Yan et al. (2025) and Kraus et al. (2024) were excluded because their operating conditions differ fundamentally. Yan et al. require source data at test time, which is explicitly unavailable in our source-free TTA setting. Kraus et al. use a distinct pre-training protocol. We will add a table enumerating these operating condition differences to situate SHOT-CCR precisely within the landscape. Figure 2 and Hyperparameters (R1, R2) The “lock” icons in Fig. 2 denote distinct stages: the backbone is frozen during initial adversarial CCR training, and the classifier is frozen during unsupervised SHOT adaptation. We will clarify this in the caption. Regarding hyperparameters, the separate learning rates and alpha were essential to strike a balance between removing noise and preserving task-relevant information. Code and Data Availability (R2) As stated in our submission, all source code and the curated JUMP-CP benchmark subset will be released publicly upon publication to ensure reproducibility. Key Summary of Contributions
- Methodological: A biologically informed, extensible, source-free TTA framework for Cell Painting image data.
- Empirical: New SOTA on RxRx1 (91.6%) and significant gains on JUMP CRISPR perturbations.
- Community: Release of a 484-class CRISPR benchmark subset.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper introduces a biologically motivated TTA method that assumes cell count as a confounder in cell painting data and suppresses this. Reviewers agreed that the idea is clear and well grounded with solid experiments.