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

Cancers exhibit phenotypically similar forms across species. Understanding conserved and diverged aspects of cancer across species can improve cancer care for both humans and animals. Transmissible cancers in dogs and Tasmanian devils offer valuable insights into immune surveillance evasion. Despite resources in companion animal care, treatment options for aggressive canine cancers are limited. Wildlife studies reveal valuable cancer models (e.g., California sea lions, raccoons). However, challenges in sample collection, data management, and analysis hinder comparative oncology. Artificial intelligence (AI) algorithms can address these challenges by enabling computational pathology. AI has revolutionized human cancer studies but lacks systematic protocols in veterinary pathology. This study aims to exploit AI's potential for pan-species tumor histology analysis by applying a human lung cancer-trained AI tool to evaluate its generalizability and accuracy in mapping tumor distribution and lymphocytic infiltration across species, and to assess the prognostic value of immune infiltration in canine melanoma and prostate carcinoma.

The literature emphasizes the value of comparative oncology in understanding cancer evolution and improving treatment strategies for both human and animal cancers. Several studies highlighted specific animal models, such as dogs and Tasmanian devils, for studying transmissible cancers and immune system interactions. The need for standardized AI protocols and digital archiving in veterinary pathology was also discussed, highlighting the gap between human and veterinary applications of AI in cancer research. Existing studies on cancer in wildlife, including those from various zoological institutions, demonstrate the importance of expanding research beyond domesticated species.

This study utilized a publicly available pan-species digital pathology atlas (pan-species.ca) comprising 120 H&E digital slide images and 41,567 single-cell annotations. A deep-learning pipeline, originally trained on human lung cancer, was applied to 99 H&E samples from 18 species from the Zoological Society of London’s archive, along with samples of Tasmanian devil facial tumor disease and canine transmissible venereal tumor. The pipeline segments viable tissue, detects cell nuclei, and classifies cells into four types: lymphocytes, stromal cells, cancer cells, and 'other' cells. The AI model’s performance was evaluated using balanced accuracy (BAcc). To assess the prognostic value of immune infiltration, 180 H&E images from canine melanoma and prostate carcinoma cohorts with clinical outcome data were analyzed. A spatial immune score, considering cell co-occurrence, was calculated using spatial statistics. A novel metric, 'morphospace overlap', was developed to quantify morphological similarity between human and non-human cells, using t-distributed stochastic neighbor embedding (t-SNE) to visualize the morphological space. Multivariate Cox proportional hazards models were used to analyze the association between immune scores and survival in the canine cohorts. An alternative AI model trained on canine prostate carcinoma was also developed for comparison.

The AI algorithm, trained on human lung cancer, achieved high accuracy (0.94 for CTVT, 0.88 for DFT1) in classifying cells in transmissible cancers. In the other 18 vertebrate species, BAcc ranged from 0.57 to 0.94, with performance correlated with cell morphological similarity to human cells. The spatial immune score, based on AI and spatial statistics, showed a significant association with prognosis in canine melanoma (HR = 0.98, p = 0.02) and a trend towards significance in canine prostate carcinoma. The 'morphospace overlap' metric demonstrated a positive correlation with the model's balanced accuracy (Pearson’s r = 0.79, p = 2 x 10-5), suggesting that morphological conservation between human and non-human cells influences the transferability of the AI model. The human-lung model outperformed a model trained on canine prostate carcinoma samples when applied to the pan-species cohort, highlighting the importance of large-scale training data and stringent quality control. Analyses of different tumour types revealed that the balanced accuracy of stromal cells varied significantly between round-cell versus epithelial and mesenchymal tumours.

This study demonstrates the potential of AI-powered computational pathology for comparative oncology, expanding its application beyond traditional model organisms. The high accuracy achieved in classifying cells in transmissible cancers and the significant association of the spatial immune score with prognosis highlight the clinical utility of this approach. The 'morphospace overlap' metric offers a valuable tool for guiding the application of AI in veterinary pathology. The study’s limitations, such as limited sample size and potential bias in species selection, should be considered when interpreting the results. Future studies should focus on expanding the dataset, improving model robustness, and validating findings across a wider range of species and tumor types. The findings underscore the importance of morphological conservation in the transferability of AI models across species and the potential for integrating AI into veterinary oncology for improved diagnostics and treatment strategies.

This study successfully demonstrated the application of AI-powered computational pathology across multiple vertebrate species. The findings highlighted the importance of morphological conservation in the successful transfer of AI models, and the clinical utility of AI-based spatial immune scoring in veterinary oncology. Future research should focus on expanding the dataset and refining the AI models to improve their accuracy and generalizability.

The study had limitations including limited sample size, potential bias in species selection, and the reliance on a model trained on human lung cancer. The generalizability of findings to all species and tumor types might be affected by these factors. Future research should address these limitations by expanding the dataset to include a more representative sample of species and tumor types.