Bridging clinic and wildlife care with AI-powered pan-species computational pathology

The study addresses how conserved and divergent aspects of cancer morphology across species can be leveraged to build AI tools for veterinary and wildlife oncology. Transmissible cancers in dogs and Tasmanian devils present unique opportunities to study immune evasion and tumor biology. Despite advances in companion animal care, treatment options for aggressive canine cancers remain limited, and systematic AI protocols for veterinary digital pathology are lacking. The authors propose that AI trained on human cancers can generalize to non-human species, enabling pan-species tumor histology analysis, spatial profiling of immune infiltration, and prognostic assessment. The purpose is to evaluate the generalizability and accuracy of a human-trained deep learning pipeline across multiple vertebrate species and to test whether AI-derived immune spatial metrics are prognostic in canine cancers, thereby advancing comparative oncology and improving animal welfare and conservation efforts.

The paper situates its work within comparative oncology, noting cancers’ phenotypic similarities across the tree of life and the potential to uncover mechanisms of tumorigenesis and resistance. Prior wildlife and zoological studies have documented neoplasia across species, including carcinosarcoma in California sea lions and papillomavirus-associated brain tumors in raccoons, as well as long-term surveys in zoos in Taiwan, France, San Diego, and Germany. While computational pathology has transformed human oncology, veterinary pathology lacks standardized AI protocols and widespread digital archiving, though efforts toward veterinary tumor classification guidelines are emerging. Previous machine learning in veterinary contexts often faced limited labeled data, prompting hybrid supervised/unsupervised approaches, and AI has been incorporated into veterinary imaging modalities. The study builds on transfer learning successes in human oncology to propose a cross-species application and addresses the gap in systematic pan-species digital pathology resources for veterinary research.

Data and cohorts: The authors curated a publicly available pan-species digital pathology atlas comprising 120 H&E whole-slide images and over 41,000 pathologist single-cell annotations for veterinary computational pathology. From the Zoological Society of London’s archive, 99 H&E samples across 18 species were digitized; additional slides included four Tasmanian devil facial tumor disease (DFT1/DFT2) and seven canine transmissible venereal tumor (CTVT) cases from the University of Cambridge. After quality control, 58 images across 20 species were retained, with one representative slide per species to form a 20-sample cohort spanning tumor types (round-cell, epithelial, mesenchymal, neuroendocrine, sex/cord stromal). For prognostic analyses, 180 canine H&E images were used: melanoma (n=88, University of Turin) and prostate carcinoma (n=12, University of Queensland) with survival data.

AI pipeline: A previously developed human-lung deep learning pipeline (predominantly lung adenocarcinoma and squamous cell carcinoma) was applied without modification to all slides. The workflow entails: (1) segmentation of viable tissue area; (2) spatially constrained CNN-based detection of nuclei centers; (3) neighboring ensemble predictor plus spatially constrained CNN classification of individual cells into four classes: cancer (malignant epithelial), lymphocytes (including plasma cells), noninflammatory stromal cells (fibroblasts/endothelial), and an ‘other’ class (e.g., macrophages, chondrocytes, normal epithelial). Performance was evaluated against expert pathologist single-cell annotations using confusion matrices and balanced accuracy (BAcc), with statistical analyses including Kruskal–Wallis and pairwise z-tests with multiple-comparison correction.

Annotations and features: Two board-certified veterinary pathologists annotated tens of thousands of cells across cohorts. Cell masks were generated with a MicroNet model pre-trained on lung H&Es, followed by automatic extraction of morphological measurements (shape descriptors, staining deconvolution, area, axis lengths, eccentricity, etc.). For interpretability, 27 features were reduced to two dimensions using t-SNE (perplexity=50, theta=0) to construct a morphological space for each class.

Morphological overlap metric: The authors developed a ‘morphospace overlap’ metric quantifying the fraction of a cell class’s 2D feature space intersecting with that of a comparator class, computed using the R package dynRB on the t-SNE projections. They correlated species-specific model performance (balanced accuracy) with mean morphospace overlap to assess transferability as a function of morphological conservation across species.

Immune spatial scoring and survival analyses: Two immune metrics were computed from AI classifications: (1) immune abundance (percentage lymphocytes) and (2) tumor–immune colocalization using the Morisita–Horn overlap index (scaled 0–100). Multivariate Cox proportional hazards models assessed associations with overall survival in canine melanoma and prostate carcinoma, adjusting for age and lymphocyte percentage and checking multicollinearity via variance inflation factor (VIF≤5). Colocalization was analyzed both as a continuous variable and dichotomized at the lower quartile. Kaplan–Meier curves were generated from adjusted models.

Digitization and QC: Slides were scanned at 40× magnification (228 nm/pixel) using NanoZoomer S210. QC excluded slides with extensive hemorrhage, necrosis, absence of tumor, or heavy pigments (e.g., melanin) that impaired cell identification. Guidelines for digitization, QC, AI execution, annotation collection, and morphospace overlap computation are provided online.

• Atlas and dataset: Released a pan-species digital pathology atlas of 120 H&E slides and over 41,000 single-cell annotations for veterinary computational pathology.
• Cross-species AI performance: The human-lung model achieved high balanced accuracy in multiple non-human species. Notably, transmissible cancers showed strong performance: CTVT in dog (overall precision 0.98; F1 and BCAcc = 0.94) and Tasmanian devil DFT1 (BCAcc = 0.88). Across 18 other vertebrate species, overall BCAcc ranged 0.57–0.94.
• Tumor-type effects: Balanced accuracy for cancer and lymphocyte classes did not differ by tumor type (Kruskal–Wallis: cancer cell H(4)=0.72, p=0.95; lymphocyte H(4)=0.534, p=0.74). Stromal-cell BCAcc differed by tumor type (Kruskal–Wallis H(4)=9.52, p=0.048), with higher stromal BCAcc in round-cell tumors compared to epithelial (z=-2.34, p=0.018) and mesenchymal (z=-2.6, p=0.02) tumors.
• Human-trained vs non-human-trained models: A model trained on non-human canine prostate carcinoma did not outperform the human-lung model on veterinary samples; e.g., for canine CTVT, BCAcc was 0.53 with the non-human-trained model vs 0.94 with the human-trained model.
• Morphological conservation and transferability: Mean morphospace overlap between animal and human cell classes positively correlated with balanced accuracy (Pearson r=0.79, p=2×10^-5), indicating species with greater morphological similarity to human training data yielded better transfer performance. Animal cancer and lymphocyte cells exhibited highest overlap with their respective human counterparts.
• Prognostic immune spatial metrics: In canine melanoma, higher tumor–immune colocalization was associated with improved overall survival (continuous HR=0.98, p=0.02, n=58; dichotomized at lower quartile: high colocalization HR=0.55, p=0.038). In canine prostate carcinoma, trends were similar but not statistically significant (continuous HR=0.95, p=0.1, n=12; dichotomized HR=0.26, p=0.13). Lymphocyte percentage alone was not prognostic in multivariate models.

The findings demonstrate that a human-trained computational pathology model can generalize to a wide array of non-human vertebrate cancers, accurately distinguishing cancer cells, lymphocytes, and stromal cells across species. This supports the hypothesis that conserved cellular morphology underpins successful transfer learning in comparative pathology. The newly proposed morphospace overlap metric provides an explainable link between morphological conservation and model performance, offering a practical guide for selecting species and tumor types most amenable to human-trained AI tools. Clinically, spatial immune colocalization—rather than immune abundance—was prognostic in canine melanoma and showed encouraging trends in canine prostate carcinoma, suggesting that spatial context of immune infiltration captures functionally relevant tumor–immune interactions in veterinary cancers, akin to observations in human oncology. These results highlight the potential for AI-enabled, cost-effective risk stratification and support integrating computational pathology into veterinary practice and wildlife care to enhance diagnostics, prognostication, and conservation medicine.

This study introduces a publicly available pan-species digital pathology atlas and demonstrates that a deep learning model trained on human lung cancer H&Es can effectively classify major cell types across diverse non-human species. The authors provide an explainable morphospace overlap metric that predicts transferability based on morphological conservation, and they show that an AI-derived tumor–immune colocalization score is prognostic in canine melanoma, with supportive trends in canine prostate carcinoma. Together, these contributions establish foundation and guidelines for transferring AI technologies to veterinary pathology and comparative oncology, with implications for precision medicine in companion animals and conservation of wildlife. Future work should expand species and tumor representation, standardize digital pathology workflows in veterinary settings, refine transfer learning strategies using the morphospace overlap metric, and validate prognostic spatial immune metrics in larger, multi-institutional cohorts.

• Limited sample sizes and annotations across species and tumor types may restrict generalizability; several cohorts, especially for survival analyses, were small (e.g., canine prostate carcinoma n=12). - Variable slide quality and a relatively low pass rate in quality control highlight differences in sample preparation and staining across institutions and species; pigments such as melanin can hinder accurate cell identification. - Transferability performance was heterogeneous across species and tumor types, with lower balanced accuracy observed in some non-mammalian samples and in stromal-class predictions for certain tumors. - The AI pipeline was trained on human lung cancers; domain shift to other tissues and species may introduce biases. - Retrospective design and reliance on archival materials may introduce selection bias; prospective validation is needed.