Source-linked AI summary
Through the Schrödinger Bridge: Benchmarking Antemortem Image Restoration from Postmortem Autolysis to Enhance Forensic Diagnostics
Shuang Hao, Jiacheng Yue, Yaxuan Zhao, Fan Wang, Jianhua Ma, Erwen Huang, Chunfeng Lian
TL;DR
The paper tackles forensic autolysis restoration, where irreversible morphological degradation and the lack of pixel-wise paired data limit conventional translation methods. It introduces AutoPath and a Schrödinger Bridge formulation, then shows that diagnostic evaluation should prioritize forensic distribution consistency over generic image-level metrics.
Problem
Irreversible, stochastic morphological degradation and unavailable pixel-wise paired samples make forensic autolysis restoration difficult for conventional supervised and structure-consistent translation methods.
Method
The paper defines unpaired autolysis restoration, introduces the homologous yet unpaired AutoPath dataset, and models restoration with a Schrödinger Bridge between autolyzed and non-autolyzed distributions.
Results
The study finds that FID and related image-level metrics do not adequately reflect diagnostic realism, whereas distribution-level consistency provides a more meaningful measure of forensic applicability.
Takeaways & Limitations
The work establishes autolysis restoration as an underexplored computational pathology task requiring task-driven, clinically grounded evaluation standards.
Takeaways & Limitations
Strict pixel-wise or structural correspondence between autolyzed and non-autolyzed images does not exist because autolysis causes irreversible biochemical and morphological changes.
Abstract
from arXiv · showhide
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 further 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.
1 Introduction
Forensic autolysis restoration addresses irreversible, stochastic tissue degradation that undermines diagnostic objectivity and defeats standard paired or structure-consistent translation assumptions. The paper establishes a benchmark combining the AutoPath dataset, a Schrödinger Bridge framework, and clinically grounded evaluation.
- Autolysis irreversibly and stochastically distorts tissue morphology, increasing diagnostic uncertainty and reliance on subjective expert judgment.The degradation includes nuclear dissolution, blurred cell boundaries, and disrupted tissue organization.
- Large, non-deterministic morphological changes and unavailable pixel-wise paired data make supervised and cycle-consistent unpaired translation methods ill-suited.Autolysis depends on temperature, humidity, and postmortem interval, with substantial variation even under similar conditions.
- The paper defines autolysis restoration as recovering severely autolyzed images toward plausible pre-autolysis morphologies under unpaired supervision.The goal is to reveal diagnostically relevant structures without assuming strict correspondence between domains.
- AutoPath is presented as the first real-world dataset for this task, using homologous yet unpaired autolyzed and non-autolyzed counterparts.The dataset contributes nearly ten thousand images from split specimen counterparts.
- The proposed framework uses a Schrödinger Bridge to model stochastic, large-magnitude morphological degradation between autolyzed and non-autolyzed domains.This formulation is intended to provide optimal stochastic transport under unpaired supervision.
- The evaluation protocol is clinically grounded because standard image-level metrics can diverge from diagnostic quality.The paper proposes slide-level diagnostic distribution consistency instead of relying only on conventional image-level metrics.
2 Method
The method constructs homologous yet unpaired autolyzed and non-autolyzed liver data, then restores images through stochastic distribution transport from the autolyzed domain to the non-autolyzed domain. It evaluates restoration through distribution-level diagnostic consistency rather than exact structural correspondence.
- Dataset construction: Each liver specimen was split into adjacent blocks, with one processed immediately and the other exposed to ambient conditions for seven days.The protocol provides medical correspondence without pixel-wise pairing because histological degradation is irreversible.
- Dataset construction: 69 patients contributed 4,962 autolyzed and 4,962 non-autolyzed training patches, with 500 patches per domain reserved for testing.Whole-slide images were scanned at 10× magnification, and 1024×1024 patches were sampled.
- Task formulation: The task learns a mapping T : X → Y from unpaired autolyzed patches to diagnostically meaningful non-autolyzed representations.Exact pixel-wise or structural preservation is not required because autolysis causes substantial cellular morphological alterations.
- Schrödinger Bridge-based Restoration Framework: The Schrödinger Bridge models restoration as optimal stochastic transport between the autolyzed distribution π0 and the non-autolyzed distribution π1.It minimizes KL divergence to a reference Wiener process while satisfying endpoint distribution constraints.
- Schrödinger Bridge-based Restoration Framework: A time-conditional generator predicts target-domain samples from intermediate states, combining adversarial distribution alignment with an entropic Schrödinger Bridge transport loss.The transport loss uses quadratic cost and entropy regularization, while structural regularization is removed to avoid preserving autolytic artifacts.
- Schrödinger Bridge-based Restoration Framework: During inference, autolyzed inputs are iteratively refined through a discretized bridge until the restored image resides in the non-autolyzed distribution.The framework also evaluates generated patches against real non-autolyzed patches using slide-level diagnostic category distributions and proportion consistency.
3 Experiments
Experiments compare the proposed model with representative unpaired translation methods on AutoPath using image-level, expert, qualitative, and slide-level evaluations. The results show that diagnostic realism and case-level consistency are better captured by expert and distributional measures than by FID alone.
- Experimental setup: The proposed method was compared with CycleGAN, CUT, DenseNorm, and KIN under identical AutoPath data splits.The evaluation protocol included image-level generative metrics, expert validation, and slide-level diagnostic assessment.
- Antemortem Image Restoration Results: The method did not achieve the lowest FID among the compared approaches.FID was used as a patch-level image distribution metric with lower values preferred.
- Antemortem Image Restoration Results: 23.81% expert preference was highest for the proposed method despite its higher FID score.Three board-certified forensic pathologists made blinded selections, and the reported percentage was averaged across experts.
- Antemortem Image Restoration Results: The Schrödinger Bridge model produced more plausible cellular morphology and tissue architecture, while competing methods often retained degradation artifacts or over-smoothed textures.The qualitative comparison focused on blurred nuclei, disrupted tissue organization, and resemblance to non-autolyzed references.
- Slide-level Diagnostic Distribution Consistency Evaluation: The method achieved the highest WSDC (0.799), lowest WKL (0.343), and highest High-Confidence Ratio (0.897) across 69 cases.Higher WSDC and High-Confidence Ratio, and lower WKL, indicate closer case-level alignment and more stable diagnostic decisions.
- Slide-level Diagnostic Distribution Consistency Evaluation: Slide-level diagnostic distribution consistency provided a more practically meaningful evaluation criterion than conventional patch-level generative metrics.Baseline methods could have competitive FID scores while showing lower weighted consistency and larger weighted divergence.
4 Conclusion
The paper establishes forensic histopathology autolysis restoration as a systematic research task and provides AutoPath alongside evaluation procedures tailored to diagnostic realism. Its conclusion emphasizes that distribution-level consistency is more informative than widely used image-level metrics in this setting.
- 4 Conclusion: The study presents the first systematic investigation of forensic histopathology autolysis restoration under irreversible, stochastic degradation and unpaired supervision.The task is positioned as an underexplored direction in computational pathology.
- 4 Conclusion: AutoPath is introduced as the first homologous yet unpaired dataset curated specifically for forensic autolysis restoration.The dataset supports investigation of restoration when pixel-wise paired samples are unavailable.
- 4 Conclusion: The proposed evaluation framework combines expert-level patch consistency with case-level diagnostic distribution consistency.This moves beyond conventional image-level metrics toward forensic applicability.
- 4 Conclusion: The experiments indicate that FID does not adequately reflect diagnostic realism, whereas distribution-level consistency offers a more meaningful measure of forensic applicability.The conclusion calls for task-driven evaluation standards in pathological image restoration.