Spatially resolved molecular profiling across complementary imaging modalities is advancing our understanding of tissue architecture, but practical integration with standard histology depends on reliable image registration to H&E sections. The authors identify common failure modes for existing registration methods: unknown image orientation, inverted image contrast, and incomplete tissue overlap. These problems can cause methods to produce biologically incorrect alignments without alerting users or attempting automated recovery.
The authors present ACCREDIT, a quality-aware, agentic registration framework that treats cross-modal alignment as an adaptive decision-making process rather than a single, one-shot computation. ACCREDIT couples deterministic registration pipelines with automated quality evaluation and an agentic rescue pathway that intervenes when registrations do not meet quality criteria.
A central component of ACCREDIT is a reference-free composite quality score that automatically evaluates registration outputs. This composite score is designed to detect biologically implausible yet superficially plausible registrations and to reject alignments that fail to meet predefined quality thresholds. The score operates without requiring an external ground-truth reference image, allowing assessment in diverse experimental contexts.
When the composite quality score indicates insufficient alignment, ACCREDIT invokes a rescue mechanism driven by a large language model (LLM)-based agent. The rescue agent autonomously diagnoses likely failure modes and selects targeted recovery strategies to correct the registration. The architecture emphasizes autonomous decision-making to identify appropriate corrective actions rather than relying on a single fallback registration algorithm.
ACCREDIT includes an optional strategy-learning module that can capture expert-validated corrections. Successful rescue strategies and expert inputs can be recorded and reused, enabling the pipeline to learn from past corrections and to improve efficiency and performance on future registration tasks.
The authors evaluated ACCREDIT across multiple cross-modal registration tasks representative of contemporary spatial-omics and histology workflows. These tasks included registration involving Xenium, CODEX, cell-boundary images, and immunohistochemistry (IHC)-to-H&E alignments. The evaluations are presented in the preprint but quantitative metrics and detailed experimental numbers are not reported in this summary beyond the qualitative comparative statements in the source.
According to the preprint, ACCREDIT outperformed competing methods by (1) detecting registration failures that other methods missed and (2) improving alignment quality through its automated recovery and rescue processes. The authors emphasize ACCREDIT’s ability to reject biologically incorrect but plausible registrations and to recover useful alignments autonomously. The source does not provide numerical performance values or specific statistical comparisons in this summary; readers should consult the preprint for full experimental results.
The ACCREDIT project repository is provided by the authors: https://github.com/LeeZhou-bearway/ACCREDIT. Supplementary materials and data/code links are referenced in the preprint. The work is presented as a preprint on bioRxiv and has not been peer reviewed.
The study lists multiple funding sources, including Department of Defense awards (HT94252410551, HT94252510959), several NIH grants (R01GM147365, R01CA283171, U01CA294548, U01CA278923, R01CA186241, S10OD034224), Cancer Prevention and Research Institute of Texas (RR260021), the Kuni Foundation, and other awards including a Knight Pilot Award (CBTOP-2023-002).
Competing interests disclosed in the preprint include financial relationships for some authors: G.B.M. (consultancies, stock/options, licensed technology, DSP patents) and R.C.S. and A.K. (consultancies and sponsored support). The remaining authors declared no competing interests. The work is released under a CC-BY 4.0 International license.
ACCREDIT is presented as a practical framework to increase robustness in cross-modal, cell-resolved image registration by combining deterministic registration methods with autonomous quality assessment and an LLM-driven rescue pathway. By rejecting biologically incorrect registrations and attempting targeted recovery, ACCREDIT aims to facilitate reliable integration of spatial molecular profiling with routine H&E-based pathology. The optional strategy-learning component offers a mechanism for accumulating expert-validated corrections to improve future performance. As a preprint, these findings should be interpreted pending peer review and users should consult the repository and full manuscript for implementation details and quantitative evaluations.