Deep learning study targets rat marrow lineages on routine H&E

A newly published Veterinary Pathology paper describes an immunohistochemistry-assisted deep learning system designed to identify cell lineages in rat bone marrow from standard H&E whole-slide images, a step that could make routine marrow review more informative in toxicologic pathology studies. The article was first published online on August 27, 2026, and lists Marco Tecilla, Edgar A. Rios Piedra, Inês B. Veiga, Kerstin Hahn, Pierre Maliver, and Smadar Shiffman among its authors. (journals.sagepub.com)

The work addresses a familiar limitation in nonclinical safety assessment. H&E marrow sections are central to evaluating cellularity and tissue architecture, but they don't always provide confident lineage-level identification on their own. In practice, pathologists may turn to bone marrow smears or immunohistochemistry for clarification, but those methods are not typically embedded in every routine workflow. According to an earlier Society of Toxicologic Pathology poster from this research group, the underlying idea was to train deep-learning models on H&E images using IHC-informed ground truth from the same tissue section, so predicted marker-expression patterns could augment standard review. (toxpath.org)

That concept also fits a broader arc in veterinary digital pathology. A 2021 review in Veterinary Pathology noted that AI-based image analysis has been moving beyond rule-based tissue quantification toward deep-learning systems that can handle more complex histology tasks across large slide sets. Separately, a 2023 paper described automated deep-learning analysis of rat bone marrow cellularity on H&E whole-slide images, showing that marrow assessment is already becoming a meaningful test case for computational toxicologic pathology. This new lineage-focused study appears to extend that progression from “how much marrow is there?” to “what kinds of marrow cells are present?” (journals.sagepub.com)

Although the full article is access-restricted, the abstract and related conference materials make the workflow clear. The team used paired H&E and IHC information from rat formalin-fixed, paraffin-embedded, decalcified sternum sections, generated from the same tissue section through sequential staining and image registration. They then trained marker-specific or lineage-oriented deep-learning models so H&E morphology could be used to predict information normally obtained through IHC. The conference abstract said those outputs could be aggregated at the cell level and translated into practical endpoints such as estimated marker expression and cell density. (visualize.jove.com)

There is precedent for this exact development strategy. In a 2025 Veterinary Pathology study on rat spleen, investigators used a destain-restain, co-registration pipeline to train a deep-learning model that inferred immune compartments directly from H&E images. That paper reported that the model outperformed unaided pathologists on certain compartment-identification tasks and argued that IHC-guided deep learning could rapidly provide quantitative information from routine sections. While spleen and marrow are different tissues, the methodological parallel is striking and suggests this marrow study is part of a larger effort to reduce dependence on extra stains for first-pass interpretation. (journals.sagepub.com)

Expert reaction specific to the new marrow paper was limited at the time of writing, but the field context is supportive. Published reviews on AI in bone marrow histology and veterinary pathology have consistently framed deep learning as a way to improve reproducibility, manage slide-scale complexity, and surface features that are difficult to quantify consistently by eye alone. At the same time, those reviews also stress familiar cautions: training-set quality, explainability, and generalizability matter, especially when models move from controlled development datasets to abnormal tissues and real-world workflows. (pubmed.ncbi.nlm.nih.gov)

Why it matters: For veterinary professionals in drug development and toxicologic pathology, the practical value isn't that AI replaces marrow interpretation. It's that AI may help extract lineage-level signals from the H&E slides already generated in standard studies, potentially improving consistency and helping pathologists decide when additional workup is needed. If validated prospectively, this kind of tool could support earlier detection of treatment-related shifts in erythroid, myeloid, or lymphoid populations, add quantitative context to narrative diagnoses, and strengthen communication with toxicologists and project teams. It may be especially useful in programs where immune-modulating or hematopoietic effects are plausible, and where quick triage of follow-up testing matters. (visualize.jove.com)

What to watch: The next milestone is likely validation in abnormal marrow and prospective use in general toxicity studies. That's where the real test lies: whether performance remains reliable when marrow architecture is altered by spontaneous disease, artifact, or test article-related injury, and whether pathologists find the outputs actionable enough to change workflow. (toxpath.org)