AI Steps Into the Oncology Clinic: How Machine Learning Is Rewriting Veterinary Cancer Care

Artificial intelligence has spent the past decade transforming human medicine, from mammography algorithms that rival radiologists to deep learning systems that screen for diabetic retinopathy in primary care clinics. Now, according to a new editorial published in the journal Veterinary Oncology, the same technological wave is breaking over animal health. Christopher J. Pinard of the Ontario Veterinary College at the University of Guelph introduces a dedicated article collection on artificial intelligence and informatics in veterinary oncology, arguing that the field has reached a genuine inflection point. Vast clinical datasets, inexpensive computational power, and a new generation of multidisciplinary collaborations are converging, he writes, to power AI-driven innovations that could redefine how cancers are diagnosed and treated in companion animals. The editorial, published as an open-access article on 7 May 2025, frames the collection as both a snapshot of current progress and a roadmap for what must happen next if these tools are to reach the clinic safely.

The technical foundations of this shift are worth unpacking. Modern veterinary AI draws on the same deep learning architectures that revolutionized human healthcare: convolutional neural networks that parse medical images pixel by pixel, natural language processing models that mine unstructured clinical records, and machine learning classifiers that integrate multi-omics data for personalized treatment planning. What has changed is not the underlying mathematics but the surrounding ecosystem. Digital pathology scanners have made whole slide images routine, picture archiving systems have accumulated years of computed tomography and magnetic resonance studies, and cloud computing has placed training-grade hardware within reach of university labs. The editorial argues that these conditions have finally aligned, allowing veterinary researchers to move from proof-of-concept demonstrations toward validated, deployable systems that can assist clinicians in real time.

One of the most striking themes in the collection is the two-way traffic between veterinary and human oncology. In human breast cancer care, radiomics-based deep learning approaches have already proven effective at predicting outcomes and have been implemented across multiple health systems, including an internationally evaluated AI system for breast cancer screening published in Nature. Veterinary researchers are now adapting those methodologies to animal patients, and, remarkably, sending innovations back across the species barrier. Bertram and colleagues demonstrated a comparative approach to mitotic count detection in which a model trained on a canine breast cancer dataset showed excellent cross-collaborative performance, supported by a completely annotated whole slide image dataset of canine mammary tumors built specifically to aid human breast cancer research.

The traffic flows in the other direction too. Osteosarcoma, an aggressive bone cancer, is far more common in dogs than in people, which means canine histopathology archives hold a wealth of data that human medicine simply cannot match. Patkar and colleagues exploited this asymmetry by training a model through deep domain adversarial learning on canine histopathology slides to detect histologic subsets of osteosarcoma in human patients. Domain adversarial learning is a technique that deliberately teaches a network to ignore species-specific features, forcing it to focus on the morphological signatures of malignancy that are shared across species. The result is a species-agnostic classifier, and the editorial points to it as evidence that comparative artificial intelligence, executed with validation and a collective domain approach, has major implications for advancing human healthcare alongside veterinary care.

Within veterinary oncology itself, the collection documents rapid progress across nearly every diagnostic modality. In digital pathology, machine learning models have achieved automated diagnosis of seven canine skin tumor types directly from H&E-stained whole slide images, while computer vision systems originally developed to detect inflammatory pododermatitis of the paw are being repurposed for neoplastic skin conditions. On the imaging side, machine learning classifiers have been trained to categorize focal splenic lesions from their CT features and to predict the histologic type and grade of canine gliomas from MRI texture analysis, offering clinicians a non-invasive preview of what a pathologist would otherwise only see under the microscope. These are not speculative demos; they are published studies with quantified performance, and they sketch a future in which pre-predicting histology before surgery becomes routine.

Perhaps the most technically consequential work concerns reproducibility, a chronic weakness in diagnostic pathology. Mitotic count, the number of dividing cells counted per high-power field, is a cornerstone of tumor grading, yet it suffers from notorious interobserver variability. Deep learning algorithms have now been shown to outperform veterinary pathologists in identifying the mitotically most active tumor region, and computer-assisted mitotic counting using a deep learning-based algorithm has been shown to improve both interobserver reproducibility and accuracy. Automated nuclear morphometry, another deep learning approach, has been applied to prognostication in canine pulmonary carcinoma. By standardizing measurements that were previously subjective, these tools attack the reproducibility problem at its root, and the editorial suggests that such techniques may become the template for AI-assisted grading across tumor types.

Data itself is the other half of the equation, and the collection highlights both progress and friction. On the progress side, the release of curated datasets such as the pan-tumor CAnine cuTaneous Cancer Histology, or CATCH, dataset is helping to standardize research and accelerate development, giving labs everywhere a common benchmark. On the friction side, an evaluation of an open-source Named Entity Recognizer applied to veterinary oncology records, conducted by Pinard and colleagues, demonstrated the potential of natural language processing to streamline data curation while exposing real limitations. General-purpose NLP systems, trained mostly on human medical text, struggled with veterinary terminology without rigorous oversight, underscoring the need for domain-specific tools. Companion reviews on quality assurance in AI model development and external validation reinforce the message that a clever algorithm is worthless, or worse, without disciplined testing.

Then comes the uncomfortable question: who is watching the deployment? The editorial is candid that veterinary medicine currently has limited to no formal guidelines for the safe and adequate deployment of AI in hospital systems. In human healthcare, concerns about data privacy, algorithmic transparency, and inherent bias have already spurred rigorous regulatory frameworks, and the risks are not hypothetical. A widely cited modelling study found that AI systems can recognize a patient’s race from medical imaging itself, a capability with troubling implications for bias. In the veterinary sphere, Duggirala and colleagues have offered a regulatory perspective outlining current initiatives and future prospects for AI and machine learning deployment, while reviews of veterinary diagnostic imaging emphasize the need for clear guidelines and safe-deployment toolboxes. The editorial argues that these initiatives will form the foundation of tailored regulatory frameworks that safeguard patient welfare and promote responsible innovation.

What makes this moment genuinely viral, in the sense of an idea spreading fast, is the realization that veterinary oncology is not merely a consumer of human AI advances but a full partner in producing them. Dogs share our environments, develop spontaneously occurring cancers with striking molecular parallels to human disease, and age fast enough to generate longitudinal data at a pace clinical trials in people cannot match. A validated model trained on canine data can seed human applications, and vice versa, creating a translational loop that benefits patients on both ends of the leash. The editorial’s closing invitation is for clinicians, researchers, and regulatory bodies to engage with the collection and to keep collaborating, because the technology is arriving whether the guidelines are ready or not.

The collection, then, captures a pivotal moment. Computer-aided diagnostics, integrated multi-omics, specialized natural language processing tools, and cross-disciplinary comparative studies are converging on a field that has historically been data-poor and resource-constrained. The editorial’s message is measured rather than utopian: AI holds real promise for diagnostic precision and therapeutic outcomes, but only if the community invests simultaneously in curated datasets, rigorous validation, domain-specific language tools, and the ethical and regulatory scaffolding that human medicine has already begun to build. If it does, the future promised in the collection’s title may arrive in veterinary clinics first, and human oncology will be better for it.

Subject of Research: Artificial intelligence and informatics applications in veterinary oncology

Article Title: The future is here: an introduction to the Veterinary Oncology collection on Artificial Intelligence and Informatics

Article References: Pinard, C. J. (2025). The future is here: an introduction to the Veterinary Oncology collection on Artificial Intelligence and Informatics. Veterinary Oncology, 2(1), Article 10. https://doi.org/10.1186/s44356-025-00027-2

Image Credits: AI Generated

DOI: 10.1186/s44356-025-00027-2

Keywords: artificial intelligence, veterinary oncology, deep learning, digital pathology, comparative oncology, mitotic count, natural language processing, radiomics, osteosarcoma, CATCH dataset, regulation, machine learning