List of Papers Browse by Subject Areas Author List
Abstract
We introduce Inverse Protocol Prediction (IPP) as the task of inferring experimental culture conditions directly from a single bright-field spheroid image. We formulate IPP as a structured multi-label prediction problem and propose a protocol-aware learning framework that integrates morphology extraction, multimodal representation learning, and dependency-aware inference. Morphometric descriptors derived from automated spheroid segmentation are fused with deep visual embeddings via a morphometry vision fusion module. To capture biological and procedural dependencies, we incorporate a Hierarchical Multi-Task Transformer (HMTT) that is conditioned to predict across protocol attributes. The framework is trained with domain-adversarial supervision and morphology-preserving augmentation and tested for cross-dataset generalisation. Hybrid convolution attention encoders achieve the best performance, reaching 95.7% multi-attribute accuracy. Results demonstrate that structured, dependency-aware modeling enables reliable reconstruction of experimental protocols from imaging alone, supporting reproducibility auditing and protocol validation in 3D cell culture systems.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5470_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/VEDAs-Lab/Inverse-Inference
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ChaJoo_Inverse_MICCAI2026,
author = { Chauhan, Joohi AND Mittal, Prateek AND Srivastava, Ayush},
title = { { Inverse Protocol Prediction from Spheroid Microscopy Imaging via Morphology-Aware Structured Learning } },
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 the task of Inverse Protocol Prediction (IPP), formulated as a structured multi-label learning problem aimed at reconstructing experimental conditions (cell line, culture medium, seeding density, and imaging setup) directly from microscopy images. Extensive experiments are conducted on the SLiMIA dataset, along with cross-dataset evaluation on RxRx1, benchmarking multiple deep learning architectures.
- 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) The paper introduces a novel problem formulation and new research direction for spheroid image analysis. Additionally, the paper demonstrates the feasibility of this direction through a comprehensive empirical study.
2) The paper introduces a custom deep learning (DL) network for mitigating acquisition bias.
- 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) The motivation for the proposed task is not adequately established in the introduction. The paper argues that protocol or metadata inconsistencies may remain undetected and affect reproducibility; however, in curated experimental datasets, metadata (cell line, media, seeding density, imaging setup) are explicitly recorded.
2) The paper lacks a dedicated methodology section, with the proposed framework primarily described through a high-level figure (Fig. 1) and scattered implementation details. Critical components such as the morphometry–vision fusion mechanism and the hierarchical multi-task transformer (HMTT) are not formally defined (no mathematical formulation, or architectural specification).
3) The paper addresses acquisition bias through a custom DL framework. However, the effectiveness of this strategy is evaluated using only a single model configuration, without comparison to alternative architectures or bias-mitigation techniques.
- Please rate the clarity and organization of this paper
Poor
- 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.
(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 paper introduces an interesting and novel research direction by formulating Inverse Protocol Prediction (IPP) for spheroid image analysis. However, the paper has several critical shortcomings that limit its suitability for acceptance (see Weaknesses). Taken together, while the paper presents a promising idea, the current formulation lacks sufficient methodological rigor, clarity, and experimental validation. Strengthening the problem motivation, providing a clear and formal methodological description, and conducting more comprehensive evaluations would be necessary for the work to be competitive.
- 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.
=> The central premise that “metadata inconsistencies undermine reproducibility in large-scale microscopy studies” is unsubstantiated. The submission neither cites a documented reproducibility failure attributable to undetected protocol inconsistencies, nor validates the workflow against a realistic auditing scenario involving synthetically or naturally mislabeled metadata. As a result, the practical utility of IPP as an auditing framework is not demonstrated.
=> The paper reports morphometry-vision fusion and dependency-aware inference as one of its primary contributions. However, the ablation shows the full model achieves Macro-F1=0.8681 versus image-only at 0.8662 (\Delta=0.0019). Consequently, the proposed components do not produce any verifiable empirical improvement over a plain image encoder.
=> Finally, the HMTT and fusion module lack formal specification; the rebuttal supplies formulations that should have appeared in the paper itself. Moreover, Figure 1, which constitutes the sole high-level description of the proposed framework, is never referenced in the text.
Review #2
- Please describe the contribution of the paper
The authors present a deep learning based pipeline for predicting experiment/imaging protocol based on microscopy images of cell spheroids.
- 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 approach is novel and addresses an existing need - there is limited prior research on the topic of inverse protocol prediction in general. In addition to the listed application areas of reproducibility auditing and protocol validation, the work has potential in data standardization in gathering large biomedical image datasets from different sources. The methods applied seem appropriate for the aims of the study, and use is generally well justified. The authors have taken steps towards reducing bias in gathering the datasets by performing ablation studies and cross-data set validation, among others.
- 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 impact of the study is limited by focusing on cross-replicate generalization rather than showing the robustness of the method – with timepoints being the only measure with held-out test classes (2 out of 109 classes). For the timepoints, a more modest prediction (F1 score of 0.75) was achieved. To my understanding, the paper does not disclose the accuracy of the model in predicting the held-out classes – this would be of interest due to being both the metric with the lowest accuracy and with held-out classes. To me, it would be intuitive to consider image intensity and contrast measures to be metrics that could inadvertently point the network to lean towards microscopy images taken in a specific study and introduce bias that way. If I have understood correctly, withholding images with a differing intensity spectrum could reveal how much the network relies on these factors instead of the textures.
- 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?
The method seems promising and addresses a novel and interesting approach to image data. There are some small shortcomings, which can be clarified without a lot of extra work.
- 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.
The answers provided by the authors have been sufficient, and with the minor clarifications suggested the by the authors in their rebuttal will further improve the paper regarding the parts where there was room for confusion.
Review #3
- Please describe the contribution of the paper
The paper introduces Inverse Protocol Prediction (IPP) for inferring experimental culture conditions directly from a single bright-field spheroid image. The authors combine automatic spheroid segmentation, morphometric feature extraction, image morphology fusion, and hierarchical multi-task prediction of protocol attributes. The paper reports very strong mean per-attribute performance and further examines biological aspects.
- 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.
Predicting protocol metadata from spheroid morphology is a well-motivated problem for reproducibility auditing and quality control in high-throughput 3D culture experiments.
The experimental split by technical replicate is a strong design choice and is more credible than random image-level splitting for this problem. It reduces the risk of obvious leakage across acquisition sessions.
The empirical study is broad with multiple segmentation models benchmarked and several IPP architectures are compared including convolutional, transformer, hybrid, fusion-based, and hierarchical variants.
The manuscript includes useful per-label analysis showing which protocol variables are strongly versus weakly recoverable from morphology.
The biological-only evaluation is a thoughtful addition because it attempts to separate biologically meaningful prediction from acquisition-driven variables.
The interpretability analysis is relevant and practically useful. The paper does not just chase high accuracy, but also tries to assess whether predictions are based on plausible biological cues versus technical confounders.
The overall practical framing is strong. Even if novelty at the module level is moderate, the task formulation and benchmarking package could be valuable.
- 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.
HMTT is positioned as a key contribution but it is not the best-performing model in the main table, and the paper does not clearly show that hierarchical conditioning improves joint protocol consistency over simpler structured baselines.
The paper should clearly explain whether uncertainty is computed across seeds and how confidence intervals were derived.
The paper would be stronger with comparison to simpler structured multi-label baselines or at least with a clearer discussion of why HMTT is preferable.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The task itself is interesting and useful. The practical motivation is strong, and the main evaluation is more careful than average because of the replicate-aware split, biological-only analysis, and per-label reporting. My impression is that the core contribution is real and that most weaknesses are addressable through better specification, clearer analysis, and tighter positioning rather than through a fundamentally different study.
- 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.
The authors have addressed all the points.
Author Feedback
We thank the reviewers and meta-reviewer for their constructive feedback and for recognizing the novelty of the IPP task, replicate-aware evaluation protocol, and biological interpretability analysis. We address the main concerns below.
Motivation and practical relevance of IPP: IPP is not intended to replace recorded metadata, but rather to serve as a complementary auditing framework for detecting protocol drift, incomplete/non-standardized annotations, acquisition inconsistencies, and cross-site variability in large microscopy workflows. Importantly, the biological-only evaluation (excluding microscope, magnification, and replicate variables) still achieves Biological Macro-F1=0.914, supporting that IPP captures biologically meaningful morphology-protocol associations rather than only acquisition artifacts.
Methodology and HMTT clarification: We agree that additional formalization would improve clarity. Due to space constraints, some implementation details were condensed. The submitted pipeline trains RefineNet for binary segmentation and freezes it before IPP training; morphometrics are computed only from predicted masks, avoiding annotation leakage. For multimodal fusion, the image encoder produces visual embeddings (z_img = f_θ(x)), while normalized morphometric descriptors (m ∈ R^d) are projected as shape tokens and fused through a Transformer encoder before per-attribute prediction heads. HMTT factorizes inference as (P(y₁,…,y_K | x) = ∏ₖ P(yₖ | x, y_<k)), using teacher forcing during training and autoregressive decoding during inference. Domain-adversarial training optimizes (L_total = L_bio+L_domain) using a gradient reversal layer. Importantly, HMTT is not intended as the sole primary contribution or necessarily the best-performing architecture. Rather, the paper systematically evaluates convolutional, transformer, hybrid, fusion-based, and dependency-aware approaches to study the role of inductive bias, explicit morphometric priors, and label-structure modeling in IPP.
Acquisition bias and robustness: Acquisition confounding is addressed through: (i) replicate-exclusive splitting; (ii) biological-only IPP excluding acquisition/replicate labels; and (iii) domain-adversarial training with replicate identity treated as domain. Grad-CAM further shows that biological predictions focus on spheroid morphology (boundary, compactness, internal organization), whereas replicate-related predictions exhibit diffuse/background attention patterns. We agree that an explicit intensity-spectrum holdout would provide an additional stress test and acknowledge this as a future direction.
Held-out classes and robustness clarification: Results are reported on held-out technical replicates rather than random image splits. For 7/8 attributes, all validation/test classes appear during training; only timepoint includes two unseen test-only classes (000h and 032h), which are included in Table 4 evaluation. The lower timepoint F1 reflects gradual and overlapping morphological transitions rather than framework failure. Robustness is further supported by zero-shot transfer to RxRx1, where Image-Shape Fusion achieves F1=0.768 despite substantial modality/acquisition shift.
Confidence intervals and reproducibility: Table 3 reports mean ± SD across three independent seeds (42,123,999), with 95% confidence intervals computed as (1.96×SD / √3). Segmentation and RxRx1 confidence intervals were computed analogously over evaluation samples. We acknowledge this could have been clearer. Code and implementation details will be released upon publication to support reproducibility.
We thank the reviewers for the thoughtful feedback. We believe the discussion clarifies that the primary contribution is the introduction of IPP as a structured inference problem together with a replicate-aware benchmarking framework for studying biological and acquisition-related morphology signatures. We will address these clarity issues in the permitted minor revision.
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.
The paper presents a method for reconstructing experimental conditions from microscopy images. As reviewers commented, it introduces a novel problem formulation. There are however various questions that lead to different ratings from the reviewers. For example, the motivation needs to be more clearly elaborated. Method descriptions should be more formal. The effectiveness for addressing acquisition bias should be more clearly demonstrated. Results on the held-out set should be included. And there are various other comments as well.
- 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.
After the rebuttal, the ratings contain one reject and two accept. The reviewer giving the reject rating has various good points, such as marginal performance improvement and unclear method formulation. Considering the method contributions of the paper, these weaknesses are relatively minor. Please make sure to improve the writing in the final version.
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.
The paper introduces Inverse Protocol Prediction (IPP) as a novel and potentially useful task for microscopy quality control and reproducibility analysis. Reviewers recognized the originality of the problem formulation, the replicate-aware evaluation design, the broad benchmarking effort, and the biological interpretability analysis. Two reviewers supported acceptance and found the rebuttal sufficient, while one reviewer remained negative mainly due to concerns about motivation, methodological clarity, and the limited empirical gain of the structured components. In my view, the rebuttal addresses the main concerns reasonably well. The authors clarified that IPP is not meant to replace recorded metadata, but to serve as an auditing and protocol-consistency tool in large microscopy workflows. They also clarified the structured role of HMTT, the morphometry-vision fusion mechanism, confidence interval computation, and the purpose of the biological-only evaluation, which still showed strong performance and supports that the method is not driven only by acquisition artifacts. The paper still has some weaknesses. The motivation could be framed more sharply, and the methodological description in the manuscript appears thinner than ideal for a MICCAI paper. In addition, the empirical advantage of some structured components over simpler baselines remains modest. However, the core contribution of the work is broader than any single module: it is the introduction and systematic study of a new structured inference task with a careful replicate-aware benchmark. Considering the novelty of the task, the practical relevance, and the generally positive reviewer response after rebuttal, I would lean Accept.
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.
The paper introduces a novel and useful problem formulation, inverse protocol prediction from spheroid microscopy images, and the overall presentation appears to have sufficient value for MICCAI.
For the camera ready version, the authors should temper claims about the proposed HMTT/fusion contribution. The reported gain of the full model over image-only appears very small. Also, the formal definitions supplied in rebuttal should be incorporated along with clearer discussion of uncertainty estimation and why the structured model is preferable to simpler baselines.
