List of Papers Browse by Subject Areas Author List
Abstract
Microscopy images provide fine-grained visual evidence of cellular morphology and are fundamental to biomedical research and clinical diagnosis. However, existing vision-language models, including CLIP and its medical variants, are primarily trained on natural images, radiology data, or medical literature, and therefore lack effective semantic alignment for microscopy imagery, limiting their performance in microscopy zero-shot recognition and cross-dataset transfer capability. In this paper, we propose MicroscopyCLIP, a domain-specific vision-language model tailored for microscopy image understanding. We construct a large-scale microscopy image-text dataset consisting of approximately 210K fluorescence microscopy images paired with textual descriptions, followed by systematic expert verification to ensure semantic consistency. Based on this dataset, we perform full-parameter fine-tuning of a pretrained CLIP model to realign the vision-language embedding space for microscopy data. Experiments demonstrate that MicroscopyCLIP substantially improves the zero-shot classification performance for microscopy images. On a held-out 8-class kidney cell benchmark, it achieves 66.17% Top-1 accuracy under prompt-based zero-shot inference without task-specific training on the target dataset. Across six external microscopy benchmarks, MicroscopyCLIP consistently outperforms existing vision-language models in both zero-shot and linear probing settings, indicating robust cross-dataset generalization. Code: https://github.com/SeriYann/MicroscopyCLIP
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1753_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanZhu_MicroscopyCLIP_MICCAI2026,
author = { Yang, Zhuoqin AND Zhang, Jiansong AND Wang, Xiaojun AND Luo, Xiaoling AND Huang, Xiaofei AND Wang, Jie AND Lim, Kian Ming AND Lu, Zheng AND Shen, Linlin},
title = { { MicroscopyCLIP: A Domain-Specific Vision-Language Model for Optical Microscopy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a microscopy-specific adaptation of CLIP for optical microscopy image understanding. Its main contribution is the combination of
- a large curated microscopy image-text corpus with 207,988 kidney cell images paired with expert-reviewed captions, and
- a two-stage progressive full-parameter fine-tuning strategy that adapts a pretrained CLIP model to microscopy semantics.
Empirically, the resulting model substantially improves zero-shot recognition on a held-out 8-class kidney benchmark (66.17% Top-1) compared with CLIP, MedCLIP, BioMedCLIP, and PMC-CLIP, and also shows strong transfer across six external microscopy classification datasets in both zero-shot and linear-probe settings.
In my view, the strongest contribution is not a new vision-language architecture, but rather a useful domain resource plus a practical adaptation recipe that appears effective for microscopy
- 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 motivation is clear and timely. The paper identifies a real gap: existing CLIP-style medical models are largely aligned to natural images, radiology, or biomedical literature, but not to fine-grained microscopy morphology.
-
The dataset construction effort is meaningful. The authors do more than generate captions automatically; they describe a hierarchical expert-review protocol and even report revision statistics for morphology, spatial relations, and identity expressions. This adds credibility to the data curation pipeline.
-
The empirical gains are strong, especially for zero-shot recognition. The jump on the held-out kidney benchmark is very large (66.17% versus 10.68%/14.94%/7.04%/9.78% for the compared baselines), and the model remains competitive across six external datasets.
-
The paper includes a useful ablation on training strategy. The two-stage progressive fine-tuning performs better than single-stage full fine-tuning and much better than LoRA under the reported setting, which supports the authors’ adaptation choice.
-
The manuscript is generally well organized. Figures 1 and 2 communicate the pipeline clearly, and the narrative is easy to follow
-
- 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 methodological novelty is limited. The model keeps the standard CLIP objective and mainly contributes data curation plus a progressive fine-tuning schedule. This makes the work more of a domain adaptation/resource paper than a clearly novel vision-language method.
-The scope of the paper is broader than what the evidence supports. The paper title and some claims suggest a general solution for optical microscopy. However, the training data appear to be concentrated on kidney cell fluorescence microscopy with only eight cell types. I would encourage the authors either to narrow the scope of the claims or to add broader pretraining data and evidence across microscopy modalities.
-
A key claimed contribution is expert verification of LLM-generated captions, but there is no direct ablation comparing raw LLM captions versus expert-verified captions. Without this, it is difficult to quantify how much of the gain comes from human verification rather than from domain-matched fine-tuning alone.
-
Although the paper positions MicroscopyCLIP as a vision-language model, evaluation is restricted to classification and linear probing. There is no retrieval or caption-alignment evaluation, even on a held-out subset of the constructed paired corpus. This weakens the evidence that the paper improves cross-modal alignment, as opposed to primarily improving visual features for classification.
-
The claim of “zero-shot” evaluation is conceptually misleading, especially for the in-domain kidney experiment. The model is trained on a kidney dataset with eight predefined cell types, and the text supervision explicitly contains class-level semantic information. It is then evaluated on a held-out split of the same 8-class kidney problem using prompts such as “A microscopy image of {class}. ” This is not zero-shot in the strict sense of recognizing previously unseen classes; rather, it is closer to label-aligned recognition without training an additional classifier head. Therefore, the statement that the model achieves “66.17% Top-1 zero-shot accuracy” is overstated and potentially misleading.
-The captioning pipeline suffers from severe label leakage. The caption-generation prompt explicitly inserts the variable “{cell_type}” and instructs the LLM to identify the cell type, justify the classification, and explain its physiological role. This means the textual supervision is not an independently observed description derived from the image alone; instead, it is conditioned on the ground-truth label from the beginning. As a result, the model may mainly learn to align images with class names and associated prior knowledge, rather than learning genuine image-text grounding. This is a major methodological flaw.
-There is a serious risk of data leakage in the claimed cross-dataset generalization experiments. The paper states that the training corpus of 207,988 microscopy images is selected from public datasets including references [22] and [9]. Later, one of the downstream external benchmarks used for evaluation is MCF7 / A549 Cytoplasm [9]. If dataset [9] contributes to both training and downstream testing, then the claim of “cross-dataset generalization” is questionable unless the authors provide a strict proof of no overlap at the image, sample, or source level. As currently written, the experimental protocol raises a serious concern about data leakage
-The main paper does not provide enough implementation detail for full reproducibility. For example, the exact CLIP backbone, precise stage-wise freezing policy, learning-rate values, batch size, augmentation strategy, stage split, dataset split sizes, and leakage-control protocol are not clearly specified in the manuscript.
-
- 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 submission does not provide sufficient information for 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?
I do not see an obvious ethics concern from the manuscript. The paper appears to use publicly available datasets and does not raise an immediate red flag regarding human-subject privacy or research integrity in the main text.
That said, the final camera-ready version should clearly document dataset licenses/usage terms and any relevant access restrictions for the released corpus.
- 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
This is a promising submission, and I believe it could become significantly stronger with a more careful experimental decomposition of the claimed contributions. In particular, I strongly encourage the authors to add: (i) an ablation of raw LLM captions versus expert-verified captions, (ii) an image-only domain-adaptation baseline trained on the same microscopy corpus, and (iii) a held-out retrieval or image-text matching evaluation on the constructed paired dataset.
I also recommend tempering the scope of the claims. As written, the paper reads as if it solves optical microscopy broadly, whereas the evidence is centered on kidney-cell microscopy and transfer to a handful of classification datasets.
Finally, please add complete implementation details and multi-seed statistics.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is a weak reject. I see clear merit in the work: the problem is relevant, the selected dataset is potentially valuable, and the empirical results are strong enough to show that the approach is useful in practice.
However, I do not think the current version is yet strong enough for acceptance as a MICCAI methodological contribution. The main reason is that the paper does not convincingly disentangle what actually drives the gains. The current evidence cannot tell whether the improvement comes primarily from (a) expert verification, (b) the mere presence of domain-matched image-text data, (c) progressive fine-tuning, or simply (d) domain adaptation of the visual encoder on kidney data.
In addition, the paper’s scope is broader than the evidence currently supports, and the cross-modal nature of the claimed contribution is under-evaluated because no retrieval or text-alignment benchmark is reported.
Overall, I find the work promising and potentially publishable after strengthening the experiments and narrowing or better substantiating the claims, but I would currently place it slightly below the acceptance threshold
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I move my recommendation from Weak Reject to Weak Accept. The authors addressed several key concerns by clarifying the expert verification process, revision ratios, LLM used, implementation details, dataset split/checkpoint selection, lack of sample-level overlap, and by agreeing to narrow the scope and use the more precise term “prompt-based zero-shot inference.”
However, one important concern remains: the caption-generation pipeline is still label-conditioned. Expert verification improves caption quality but does not fully remove the risk that the model learns class-name alignment rather than genuinely image-grounded vision-language correspondence. The raw-versus-verified caption ablation and new retrieval results are useful, but they should be clearly reported and interpreted cautiously.
Overall, I view the paper as a valuable microscopy-domain resource and practical CLIP adaptation study, rather than a strong methodological contribution. With the clarified scope, implementation details, no-overlap documentation, retrieval results, and caption ablation included in the camera-ready version, the paper is acceptable for MICCAI.
Review #2
- Please describe the contribution of the paper
This paper presents MicroscopyCLIP, a domain-specific vision-language model for optical microscopy. The main contribution is the construction of a microscopy image-text pair of about 208K kidney cell images. Text captions are generated by an LLM but revised by experts. A ViT based model was trained using two-stage domain adaptive fine-tuning. The model is evaluated on an in-domain held-out kidney dataset and six external microscopy benchmarks under zero-shot and linear-probe settings. The main contribution of the paper is creating a large dataset and the empirical gains of domain specific training over CLIP, MedCLIP, BioMedCLIP, and PMC-CLIP for zero-shot microscopy classification.
- 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.
Large microscopy dataset The paper identifies a very important gap in the field. Existing vision-language models in biomedicine are indeed dominated by radiology, pathology, and general biomedical image-text pairs, while fine-grained optical microscopy remains underrepresented. Authors collected 208k cell images from kidney tissue spanning 8 different cell types.
Expert revised captioning The captions were first generated by an LLM in a structured way. Then the hierarchical expert revisions were conducted. Junior researchers did the first revision, followed by the senior expert revision.
Training strategy The benefit of two-stage training strategy is well demonstrated by ablations. Two-stage alignment shows better performance than one-stage full-layer training.
- 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.
Lack of novelty The paper does not present a methodological contribution. However, I would not consider this a major concern, as the primary motivation of the work is to develop a model that produces better embeddings for microscopy images.
Dataset splits are not defined Although the training dataset (208k images) is described in detail in Figure 1, the held-out test set is not clearly defined. It would be helpful to report the number of cells in each class in the test set.
Scope of the dataset Although the image-text pairs fill a very important gap, the dataset consists only of human kidney tissue spanning eight different cell types. In future studies, it could be improved by including additional tissues in the training dataset.
LLM generated captions are (too) structured Because the captions appear to follow a highly consistent structure, the model may partially rely on this fixed textual format during alignment. Introducing greater variation in phrasing or in the ordering of descriptive components could encourage the model to learn more robust image-text associations rather than patterns tied to caption structure alone. Such variation could also function as a form of text augmentation.
Evaluations are narrow Evaluations are limited to zero-shot classification and linear probing. As acknowledged in the manuscript, “Since the downstream microscopy benchmarks do not provide paired image-text annotations, cross-modal retrieval evaluation is not applicable in this study.” However, additional analyses such as image-image retrieval, UMAP visualizations, or other embedding-quality metrics, including kNN-based evaluation, could still be included. A strong visualization would be showing the embeddings before the alignment and after the alignment in one of the external cohorts.
- 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
Which LLM was used to generate reports?
How many generated captions were reviewed by experts? Was all of the data manually reviewed by experts, or did you perform expert review only on a randomly sampled subset, followed by further refinement or correction using the LLM?
Could you clarify this section: “The revision rates exhibit a hierarchical gradient: 47.5% at the morphological level (D2), largely involving refinement of nuclear morphology descriptions and removal of imprecise or overly generalized visual statements; 66.3% at the spatial level (D3), clarifying structural localization or eliminating unsupported spatial inferences; and 17.6% at the identity level (D4), mostly consisting of terminology normalization rather than category changes.” What do these ratios represent? Fraction of generated reports needed to be corrected in each category? Fraction of generated reports reviewed by the experts?
How was the best checkpoint selected? Was it watched over the internal held-out dataset?
- 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 contributes an image-text paired dataset to the microscopy field.
The model achieves better performance compared with other biomedical CLIP models.
The evaluation includes only zero-shot classification and linear probing, and could be further enriched.
It is unclear how many generated reports were reviewed by experts, which raises questions about the quality of the dataset.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Although novelty and the scope of the paper is limited, the effort for cell-caption pairs are significant. My concern was validity of captions created for images as they were initially created with an LLM. Authors now clarify that all captions went under review by experts. Authors also acknowledge that the scope of the paper is limited to kidney and they will tune down the claims for all microscopy.
Review #3
- Please describe the contribution of the paper
The paper presents MicroscopyCLIP, a microscopy-domain adaptation of CLIP for optical microscopy image understanding. Its main contribution is the construction of a microscopy image-text corpus of roughly 208K kidney cell images with LLM-generated captions that are subsequently manually revised and expert-verified, followed by domain-adaptive fine-tuning of a pretrained CLIP model using the standard symmetric contrastive objective. The paper further studies a two-stage progressive fine-tuning strategy, in which later layers are adapted first before full end-to-end tuning, and demonstrates improved zero-shot recognition and cross-dataset transfer on multiple microscopy benchmarks.
- 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.
A main strength of the paper is its experimental evaluation. The authors test on many datasets, covering both in-domain and cross-dataset settings, compare against several relevant baselines, and include ablation studies on the fine-tuning strategy and prompt design. Overall, the empirical section is fairly comprehensive.
- 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 has several weaknesses. First, although the authors emphasize the construction of a large-scale microscopy image-text dataset, the paper does not provide enough detail about how the captions are actually generated, including the specific prompt design principles, the generation rules, and the rationale behind the final caption formulation. The expert verification process is also insufficiently documented: the paper does not clearly explain how the review was conducted in practice, how many samples were examined, or provide concrete examples of revision cases, which makes the reliability of the caption quality control less convincing. Second, the technical depth of the paper appears limited. The overall framework mainly combines LLM-based caption generation, standard CLIP-style contrastive training, and a simple two-stage fine-tuning strategy, without demonstrating a clearly novel model design or training objective tailored to microscopy. Finally, the paper lacks important implementation and reproducibility details, such as batch size, learning rate, and other training settings, making it difficult to fully assess or reproduce the reported results.
- 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 submission does not provide sufficient information for 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main reason for my recommendation is that the paper lacks sufficient technical depth and novelty. The overall method largely consists of LLM-based caption generation, standard CLIP-style contrastive training, and a simple two-stage fine-tuning strategy, without a clearly distinctive modeling or optimization contribution tailored to microscopy. In addition, the paper does not provide enough experimental and implementation details to support reproducibility, such as batch size, learning rate, and other key training settings. The dataset construction and caption verification pipeline are also insufficiently described, especially with respect to prompt design, expert review procedure, and quality control details. Although the experimental section is relatively broad, the limited technical contribution and the lack of reproducibility details make the overall paper unconvincing in its current form.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After reading the authors’ rebuttal, my concerns were not sufficiently addressed. In particular, the authors did not fully respond to my questions. The method still appears to lack fundamental technical novelty, as it mainly combines LLM-based caption generation, standard CLIP-style contrastive training, and a relatively simple fine-tuning strategy without a clearly microscopy-specific model design or objective. In addition, the paper still does not provide enough detail about intermediate data processing, caption construction, expert verification, or data augmentation procedures. These missing implementation details make the work difficult to reproduce and weaken confidence in the reported results. Therefore, I maintain my decision to reject the paper.
Author Feedback
We sincerely thank all reviewers for their constructive feedback. We address the major concerns below.
[R1] Label leakage in the captioning pipeline
We agree that the caption generation process is label-conditioned. However, the label is not used as the sole supervision target, but as contextual guidance for generating structured biological descriptions. Importantly, expert reviewers subsequently verified and revised captions against the underlying microscopy image content, particularly for morphological and spatial descriptions. Therefore, the final supervision signal is not limited to class identity alone, but also contains image-grounded biological information.
[R1] Data leakage
There is no sample-level overlap. Reference [9] is a paper aggregating multiple independent image collections. Our training data draws exclusively from the kidney cortex fluorescence microscopy subset. MCF7/A549 Cytoplasm is a completely separate subset with different cell types.
[R1, R2, R3] Methodological novelty
We agree that the primary contribution of MicroscopyCLIP is not a fundamentally new VL architecture, but rather a microscopy-specific VL resource and domain-adaptive alignment framework. Existing biomedical CLIP-style models mainly focus on radiology, pathology, or biomedical literature, while microscopy-specific vision-language alignment remains largely underexplored. To our knowledge, MicroscopyCLIP is among the first large-scale VL alignment efforts specifically targeting this underexplored gap.
[R1] “Zero-shot” terminology
We acknowledge that “zero-shot” may be interpreted differently in the broader ML literature. Our usage follows the original CLIP definition. We will also revise the manuscript to use the more precise term “prompt-based zero-shot inference”.
[R1, R2] Raw LLM captions vs. expert-verified captions
Initial captions were generated using GPT-4o, and all captions underwent manual expert verification. The revision ratios (e.g., D2: 47.5%) represent the percentage of captions requiring manual correction in the corresponding category. To further evaluate the impact of expert verification, we compared raw and verified captions. Raw generated captions achieve 55.89% (Kidney) and 25.23% (Jurkat), while expert-verified captions achieve 66.17% and 32.26%, confirming consistent gains from expert verification.
[R1, R2] Retrieval evaluation
We agree retrieval evaluation strengthens the paper. We therefore generated paired text descriptions for the held-out test set using the same captioning pipeline and conducted image-text retrieval evaluation. Specifically, image-to-text retrieval improves from 3.1%/6.1% (CLIP) and 1.8%/4.5% (MedCLIP) to 12.4%/22.5% at R@5/R@10, while text-to-image retrieval improves from 2.7%/5.5% (CLIP) and 2.4%/5.1% (MedCLIP) to 9.4%/18.9%. MicroscopyCLIP consistently improves retrieval performance over both baselines. Additional retrieval metrics and UMAP visualizations will be included in the camera-ready version.
[R1, R2, R3] Implementation details
Here we provide the implementation details and will include a complete table in the camera-ready version. Backbone: ViT-B/32 initialized from OpenAI CLIP. Effective batch size: 1024 across 4 GPUs with mixed-precision training. Stage I: lr = 1e-4 for 5 epochs. Stage II: lr = 1e-5 for 25 epochs. Training and validation subsets were created using stratified 85%/15% splitting within each class. The held-out kidney test set consists of a separately predefined subset containing 47,280 images. Exact split statistics and per-class sample counts will be included in the camera-ready version. The best checkpoint was selected based on validation loss.
[R1, R2] Scope of claims
We accept this criticism and will narrow the title and abstract to more precisely reflect the current scope: fluorescence microscopy of human kidney tissue, with demonstrated transfer to 6 external datasets. Broader microscopy modalities will be explored in future work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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 presents a new image-text dataset on kidney cell images which is used to fine tune a pretrained CLIP model. Reviewers agree with the two contributions of the paper: the dataset and fine-tuning strategy. There are a list of comments and concerns, leading to mixed review scores (Weak Reject, Weak Accept, and Reject). Some major weakness items include: the methodology is lack of novelties; the scope is over-claimed, the expert verification on image captions is unclear, labeling and experiment setting are questionable, and the technical implementation is missing.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After the rebuttal, all reviewers still have a long list of concerns in addition to the previous critical concerns such as the lack of novelties and limited scope. After reading the paper, rebuttals and reviews after the rebuttal, the AC recommends rejecting this draft this time.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper contributes a fluorescence microscopy image-text dataset on human kidney tissue and a domain-adaptive CLIP fine-tuning framework. The rebuttal addressed the main concerns well—adding image-text retrieval results, quantifying the gain from expert-verified captions, ruling out data leakage, and providing implementation details. The authors also appropriately narrowed their claims and the “zero-shot” terminology. The remaining concern about limited technical novelty is valid, but the contribution is reasonably positioned as a domain-specific resource and adaptation framework, which fits MICCAI’s scope.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This masnucript introduces a large image-text kidney cell image dataset and a practical fine-tuning strategy for CLIP-style model training. Although this work may not exhibit strong technical novelty, the proposed dataset and the CLIP adaptation method are valuable to the biomedical image analysis community. In addition, the rebuttal addresses several key concerns from reviewers, such as clarification of image caption verification, potential data leakage, implementation details, and scope of claims.