List of Papers Browse by Subject Areas Author List
Abstract
Forensic histopathology, essential for determining cause of death and disease diagnosis, is severely impeded by postmortem autolysis, i.e., an irreversible, stochastic degradation process that distorts tissue morphology and introduces diagnostic subjectivity, thereby underscoring the value of restoring autolyzed images to a diagnostically plausible, pre-autolysis state for improving objectivity in forensic practice. This restoration task is fundamentally challenging due to the large, non-deterministic morphological changes caused by autolysis and the infeasibility of pixel-wise paired data, which invalidates assumptions underlying supervised and cycle/structure-consistent unpaired translation methods. To address this, we formalize forensic histopathology autolysis restoration as a new task: under unpaired supervision, transform postmortem images with severe autolysis into diagnostically meaningful ``antemortem’’ representations. We contribute AutoPath, the first homologous yet unpaired dataset for this problem, constructed by splitting specimens into adjacent tissue blocks—one processed immediately, the other exposed to induce autolysis—yielding nearly ten thousand $10\times$ patches from 69 cases with varying liver conditions. We fruther frame the problem as a Schrödinger Bridge between the autolyzed and non-autolyzed distributions, offering a principled approach to modeling stochastic, severe morphological degradation. Critically, we demonstrate the misalignment of generic image-level generative metrics (e.g., FID) with diagnostic utility and propose a forensically grounded, slide-level diagnostic distribution consistency evaluation. Overall, this work establishes a reproducible benchmark (encompassing task definition, a real-world dataset, and an evaluation methodology) toward rigorous and practically meaningful progress in autolysis restoration for forensic pathology.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0259_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/0259_supp.zip
Link to the Code Repository
https://github.com/fengluo233/AutolysisBench
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HaoShu_Through_MICCAI2026,
author = { Hao, Shuang AND Yue, Jiacheng AND Zhao, Yaxuan AND Wang, Fan AND Ma, Jianhua AND Huang, Erwen AND Lian, Chunfeng},
title = { { Through the Schrödinger Bridge: Benchmarking Antemortem Image Restoration from Postmortem Autolysis to Enhance Forensic Diagnostics } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a new and clinically meaningful task in computational pathology: restoring severely autolyzed postmortem histopathology images toward diagnostically plausible antemortem representations under unpaired supervision.
The authors make three key contributions:
Task and Dataset: They formalize forensic autolysis restoration as a new problem and introduce AutoPath, the first homologous yet unpaired dataset constructed by splitting adjacent tissue blocks from the same specimen, enabling medically meaningful supervision despite the absence of pixel-wise pairing. Methodology: They formulate the problem as a Schrödinger Bridge (SB) stochastic transport between autolyzed and non-autolyzed distributions, which is conceptually well-aligned with the irreversible and stochastic nature of autolysis. The model avoids structural consistency constraints and instead focuses on distribution-level alignment. Evaluation Framework: They propose a slide-level diagnostic distribution consistency metric, demonstrating that commonly used generative metrics (e. g. , FID) are misaligned with forensic diagnostic utility, and advocate for clinically grounded evaluation.
Overall, the paper establishes a new benchmark for this underexplored problem and provides a framework for evaluating restoration quality in a clinically meaningful way.
- 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.Novel and clinically relevant problem formulation The paper addresses an important and underexplored challenge in forensic histopathology. The distinction between autolysis restoration and conventional image translation is clearly articulated and well-motivated. 2.Well-designed dataset with strong practical grounding The AutoPath dataset is a significant contribution. The use of spatially adjacent tissue blocks to construct homologous yet unpaired samples is a thoughtful and realistic solution to the lack of paired data. 3.Conceptually appropriate modeling choice The use of Schrödinger Bridge to model stochastic, irreversible degradation is well aligned with the problem characteristics. The decision to remove structural consistency constraints is also justified and differentiates this work from standard unpaired translation methods. 4.Clinically meaningful evaluation beyond FID The paper convincingly shows that lower FID does not correspond to better diagnostic quality. The proposed slide-level diagnostic distribution consistency is a strong contribution and more aligned with real-world forensic practice. 5.Comprehensive evaluation (quantitative + human + case-level) The combination of patch-level metrics, expert validation, and slide-level consistency provides a more holistic evaluation than typical generative modeling papers.
- 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 methodological novelty beyond adaptation of SB framework While the formulation is well motivated, the model itself appears to be an adaptation of existing neural Schrödinger Bridge approaches with adversarial alignment. The primary novelty lies more in task formulation and evaluation rather than algorithmic innovation. 2.Insufficient implementation details and reproducibility concerns Key aspects of the method are under-specified, including model architecture, training schedule, SB loss implementation, and hyperparameters. This limits reproducibility. 3.Evaluation relies on a potentially weak intermediate classifier The slide-level evaluation depends on a classifier trained with ~300 annotated patches. The robustness and calibration of this classifier are not sufficiently validated, raising concerns about the reliability of WSDC/WKL metrics. 4.Limited statistical rigor in expert evaluation The expert validation results are promising but lack statistical analysis (e. g. , inter-rater agreement, confidence intervals, significance testing). The evaluation protocol may also introduce forced-choice bias. 5.Risk of hallucination and clinical misinterpretation not fully addressed Since the task does not enforce structural fidelity, the model may generate plausible but incorrect morphology. The paper does not sufficiently discuss the potential risks of such hallucinations in forensic diagnosis. 6.Limited dataset scope and generalization The dataset is restricted to liver tissue and a relatively small cohort (69 cases). It is unclear how well the method generalizes to other organs, magnifications, or varying degrees of autolysis. Baselines may not be sufficiently strong or recent 7.Comparisons are mainly against GAN-based unpaired methods. Stronger diffusion-based or medical-specific baselines would strengthen the claims.
- 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 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?
The proposed method raises potential concerns regarding hallucination of non-existent pathological structures. Since the model aims to generate diagnostically plausible “restored” images without enforcing structural consistency, it may introduce artificial morphology that does not correspond to the true underlying tissue.
In forensic settings, such hallucinated features could potentially mislead diagnosis or interpretation. The paper would benefit from a clearer discussion of intended use (e. g. , visualization vs. diagnostic support) and safeguards against misuse.
- 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 an interesting and well-motivated work addressing a clinically meaningful problem. I particularly appreciate the effort to go beyond standard generative metrics and propose a more domain-relevant evaluation framework.
To further strengthen the paper, I suggest:
Providing more detailed methodological descriptions to improve reproducibility. Adding statistical analysis to expert validation (e. g. , agreement metrics, significance testing). Evaluating the robustness and calibration of the slide-level classifier. Including failure cases and analyzing when the model produces unrealistic or misleading structures. Expanding discussion on clinical applicability and risks of hallucination. Considering stronger or more recent baselines.
Overall, the work has strong potential as a benchmark and problem-defining contribution.
- 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 novel and clinically relevant problem and introduces a valuable dataset and evaluation framework. The experimental results provide useful insights, particularly regarding the limitations of standard generative metrics in medical contexts.
However, the methodological novelty is moderate, and there are concerns regarding reproducibility, evaluation rigor, and potential clinical risks (e.g., hallucination). The work is still promising and likely to stimulate further research in this area, but some aspects require clarification and strengthening.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
This paper proposes a highly novel computational pathology task: utilizing a deep generative model (Schrödinger Bridge) to restore severely autolyzed H&E-stained slides to a state closely resembling fresh tissue. To overcome the physical impossibility of obtaining pixel-level ground truth, the authors cleverly construct a “homologous yet unpaired” liver tissue dataset. Overall, the paper has an excellent entry point, clear motivation, and holds significant practical value for forensic and clinical pathology. However, the manuscript currently exhibits noticeable shortcomings in the comprehensiveness of its quantitative evaluation, the validation of its downstream clinical utility, and its cross-organ generalizability, which require further improvement.
- 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.High Novelty and Forensic/Clinical Significance: Unlike previous “bypass” strategies that rely on specialized equipment (e.g., virtual staining via autofluorescence microscopy), this paper tackles the most challenging and common issue head-on: restoring routine H&E slides that have suffered from autolytic degradation. The introduction of this task fills a critical gap in the field and has immense practical application value. 2.Clever Dataset Design: The “homologous yet unpaired” data collection strategy elegantly addresses the impossibility of acquiring paired Ground Truth due to the destructive nature of tissue preparation. This provides an excellent and rigorous experimental paradigm for future research in this domain. 3.Reasonable Methodological Choice: Given that autolysis inherently destroys cellular structures, traditional unpaired translation methods relying on cycle consistency (e.g., CycleGAN) are fundamentally unsuitable. The authors’ choice to utilize the Schrödinger Bridge to model the stochastic dynamics between complex, structurally disparate distributions is logically sound and perfectly tailored to the task’s characteristics.
- 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.Insufficient Quantitative Metrics: The current evaluation lacks standard distribution-level metrics for generative models (e.g., FID, KID) to objectively measure image quality. Furthermore, a subjective evaluation by professional pathologists (e.g., Mean Opinion Score or a Turing Test) is missing, which is crucial for assessing morphological fidelity and the presence of generative artifacts. 2.Lack of Downstream Validation: The paper does not evaluate whether the restored images actually recover diagnostic value. Including a standard computational pathology task—such as nuclei instance segmentation or lesion classification—on the restored images is essential to prove the practical and clinical utility of the proposed method. 3.Limited Scope (Liver Tissue Only): The experiments are exclusively conducted on liver tissue. Since different organs (e.g., kidney, brain, heart) exhibit vastly different autolysis mechanisms and morphological changes, the generalizability of the proposed framework to other tissue types remains unverified.
- 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 conceptualization of this paper is outstanding, but the depth and breadth of the experiments do not yet fully support its ambitious task setting. If the authors can supplement the Rebuttal/Revision phase with downstream task validation and richer quantitative metrics, this will be a highly impactful and excellent piece of work.
- 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 #3
- Please describe the contribution of the paper
This paper introduces a new task of forensic histopathology autolysis restoration, aiming to transform postmortem autolyzed tissue images into diagnostically plausible antemortem representations under unpaired supervision. The authors contribute a novel dataset (AutoPath) constructed using adjacent tissue blocks to approximate correspondence between autolyzed and non-autolyzed samples, and formulate the restoration problem as a Schrödinger Bridge to model stochastic morphological degradation. Additionally, the paper proposes a clinically motivated evaluation framework based on slide-level diagnostic distribution consistency, highlighting limitations of conventional generative metrics such as FID.
- 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.
Novel problem formulation: The paper defines an important and underexplored task in computational pathology—autolysis restoration for forensic diagnostics—which is both clinically relevant and technically challenging.
Dataset contribution: The AutoPath dataset is a valuable contribution. The use of adjacent tissue blocks to approximate homologous but unpaired samples is a thoughtful and practical design that reflects real-world constraints.
Principled modeling approach: Framing the problem as a Schrödinger Bridge is theoretically well-motivated for modeling stochastic, irreversible morphological degradation. This is a strong methodological choice for unpaired translation with large domain shifts.
Clinically aligned evaluation: The paper correctly identifies limitations of FID and introduces a slide-level diagnostic distribution consistency framework, which is more meaningful for clinical applications.
Inclusion of expert evaluation: The use of board-certified forensic pathologists for blinded evaluation adds credibility and relevance to the results.
Comprehensive evaluation pipeline: The combination of patch-level metrics, expert validation, and slide-level consistency provides a multi-faceted evaluation strategy.
Overall, the combination of a new task, dataset, and clinically motivated evaluation makes this a meaningful contribution to the field.
- 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.
Limited dataset scale and diversity: The dataset consists of 69 cases from a single organ (liver), which may limit generalizability across tissues, disease types, and forensic scenarios.
Weak supervision and label dependency in evaluation: The slide-level evaluation relies on a classifier trained on a relatively small number of annotated patches (~300), which may introduce bias and affect the reliability of the diagnostic distribution consistency metric.
Lack of strong quantitative improvement: The method does not achieve the best FID score and the improvement in expert preference (23.81%) is relatively modest, raising questions about the magnitude of practical gains.
Ambiguity in restoration validity: Since autolysis is irreversible, there is no ground truth for restored images. While the authors acknowledge this, the framework still assumes that generated images are diagnostically meaningful without rigorous validation beyond preference-based evaluation.
Risk of hallucination in generative model: The Schrödinger Bridge model may introduce plausible but incorrect morphological structures, which could mislead downstream diagnosis. This risk is not deeply analyzed.
Evaluation depends on downstream classifier: Slide-level consistency metrics rely on a trained classifier, making the evaluation indirect and dependent on classifier performance rather than purely on restoration quality.
More importantly, the absence of ground truth for restoration combined with reliance on indirect evaluation metrics makes it difficult to assess whether the generated images are diagnostically correct or merely visually plausible.
- 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
This paper presents an interesting and important direction for forensic pathology, combining dataset creation, modeling, and evaluation. However, stronger validation of restoration correctness and broader evaluation across datasets would improve confidence in the approach. Additionally, clearer discussion of ethical considerations and potential risks of generative hallucination would strengthen the work.
- 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?
This paper introduces a novel and clinically meaningful problem in forensic histopathology and provides a well-designed dataset and principled modeling approach. The use of a Schrödinger Bridge for modeling stochastic degradation is technically sound, and the evaluation framework appropriately emphasizes clinical relevance over traditional generative metrics.
However, several limitations reduce confidence in the results. The dataset is relatively small and limited in scope, and the evaluation relies on indirect measures such as expert preference and classifier-based distribution consistency. Additionally, the inherent ambiguity of the restoration task, combined with the risk of hallucinated structures, raises concerns about whether the generated images are clinically reliable or potentially misleading.
Overall, the paper is slightly above the acceptance threshold due to its novelty and practical importance, but would benefit from stronger validation and broader evaluation.
- 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 all reviewers for their valuable comments and positive feedback. We are encouraged that reviewers consistently recognized the novelty and forensic significance of the task, the value of the AutoPath dataset, the suitability of the Schrödinger Bridge (SB) formulation, and the importance of clinically grounded evaluation beyond FID. We emphasize that this work is primarily intended as a benchmark-defining and clinically grounded formulation for forensic autolysis restoration, rather than a purely architecture-driven generative modeling paper. R2Q1&R2Q7: Method novelty and baselines The main novelty lies in: (1) formalizing forensic autolysis restoration as a new computational pathology task (2) constructing the first homologous-yet-unpaired benchmark (AutoPath) (3) introducing clinically grounded evaluation beyond conventional perceptual metrics. SB is adopted as a principled framework for modeling irreversible and stochastic autolysis. Since severe autolysis physically destroys tissue structures, strong structural consistency assumptions are unsuitable for this task. Thus, we focus on distribution-level restoration rather than architectural novelty alone. We used representative unpaired translation methods as baselines and agree that stronger diffusion/pathology-specific baselines would be valuable future extensions. R2Q2: Reproducibility In the camera-ready version, we will further provide implementation details, training settings, and hyperparameters. We also plan to publicly release the code and AutoPath dataset. R2Q3&R5Q2&R5Q6: Reliability of the classifier The classifier is intended as a proxy for slide-level diagnostic distribution consistency rather than an absolute diagnostic oracle. We adopt CONCH pretrained weights, which are extensively validated in computational pathology. Although only ~300 annotated patches were used, we only fine-tuned a lightweight MLP head while keeping the encoder fixed, resulting in stable convergence. The downstream categories also correspond to relatively standard histopathological patterns. R2Q4&R3Q1&R5Q3: Evaluation and quantitative metrics A key motivation of this work is to demonstrate that conventional perceptual metrics such as FID may not align with clinically meaningful restoration quality. Besides FID, we additionally included blinded expert evaluation and slide-level diagnostic consistency evaluation. The discrepancy between FID and slide-/expert-level evaluations further supports the need for clinically grounded evaluation. R2Q5&R5Q4&R5Q5&R5Q7: Hallucination and restoration validity Our goal is not pixel-wise reconstruction of the original tissue, which is physically impossible under irreversible autolysis and destructive tissue processing. Instead, we aim to restore diagnostically plausible distributional representations. This is also why we avoid enforcing strict structural consistency constraints. Under severe autolysis, forcing strong structural alignment often produces anatomically implausible artifacts because many cellular structures are irreversibly destroyed. To address the lack of direct ground truth, we employ clinically grounded indirect evaluations, including patch-level analysis, slide-level diagnostic consistency, and expert assessment. We position this framework as a visualization and decision-support tool under expert supervision rather than an autonomous forensic diagnostic system. R2Q6&R3Q3&R5Q1: Dataset scope and downstream validation Liver tissue is one of the most important organs routinely sampled in forensic autopsy and is particularly suitable for studying autolytic degradation. We agree that broader multi-organ validation would further strengthen the benchmark and are currently expanding the dataset scale and diversity. We also agree that downstream validation is important. Our slide-level diagnostic consistency evaluation itself serves as a clinically motivated downstream task focuing on disease distribution consistency after restoration.
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.
The paper’s main strengths are its novel and clinically important formulation of autolysis restoration in forensic histopathology, the carefully designed AutoPath dataset using homologous yet unpaired adjacent tissue blocks, and the well-motivated use of Schrödinger Bridge to model stochastic and irreversible degradation. The evaluation is also clinically meaningful, combining patch-level metrics, expert assessment, and slide-level diagnostic consistency beyond standard FID.
However, the reviewers also noted several weaknesses. The model has limited methodological novelty, as it largely adapts existing Schrödinger Bridge approaches, and the dataset is restricted to a relatively small liver cohort, leaving generalization unclear. More importantly, the evaluation has key limitations: the slide-level metrics rely on a weakly validated classifier, expert evaluation lacks stronger statistical analysis, and there is no ground truth to confirm whether the restored images are diagnostically correct rather than merely visually plausible. The paper also insufficiently addresses hallucination risks, downstream diagnostic validation, and comparisons with stronger recent baselines.
Having said this, I find that the paper balances between strong and weak aspects, leaning towards the positive side. I believe the main strength lies in its strong clinical usefulness for restoring degraded slides, an important problem that has not yet been actively studied. I also note that all three reviewers unanimously agree that the paper is acceptable and highly praise its clinical usefulness and practical value. Based on this assessment, I recommend acceptance.
