AI and the Diagnostic Sample: From Cytology Reads to Smarter Biopsy Selection
AI in Veterinary Diagnostics · Part 2 of 6
I didn't know veterinary pathology existed. That sounds like an odd thing to admit, but it's true. The summer before my first year of veterinary school, I found myself in a pathology department for the first time — and I was stopped in my tracks. There were pathologists who could look at a glass slide, nothing else, and tell you not just what tissue they were looking at but what disease was present. No hints. No history. Just tissue and knowledge.
That summer I also read The Coming Plague by Laurie Garrett — a book about emerging infectious diseases and the detectives who track them down. Pathology and epidemiology suddenly felt like the most important work in medicine. I was hooked on the idea that disease had a language, and that some people had learned to read it.
It still took a few years in private practice before I fully accepted that pathology was my calling. But that summer planted the seed. And the thing that struck me most — then and now — is how much a good diagnosis depends on the quality of the sample behind it. What gets collected. How it's documented. Whether it gets submitted at all.
Where diagnostic information gets lost
The diagnostic sample is where a clinical impression becomes something testable. It's also where a surprising amount of information disappears — through poor sampling, inadequate fixation, or simply the decision not to submit. These aren't lab failures. They happen upstream, in the clinic, before the sample ever leaves.
AI is beginning to change that — and the implications are worth thinking through carefully. Some of what's emerging is genuinely useful. Some of it is being oversold. The challenge, as with all AI tools in veterinary medicine, is knowing the difference.
What AI can and can't read in cytology
Cytology is fast, affordable, and widely available — and it's also highly dependent on the person collecting and reading the sample. The same fine needle aspirate can tell very different stories depending on how it was prepared and who interprets it.
AI-assisted cytology has made real progress in human medicine, particularly in cervical screening and blood cell analysis, where large standardized datasets made model training possible. The translation to veterinary cytology is harder. Veterinary cytology spans dozens of species, huge variation in sample preparation, and far smaller training datasets. The bar for meaningful AI validation is much higher as a result.
Context from human medicine
AI-assisted cervical cytology screening has been in use in human medicine for decades, validated in large studies. The key was standardized preparation, high case volumes, and a limited range of cell types to classify. Veterinary cytology has none of those advantages at scale — which is why the validation work needed here is genuinely more complex.
What is showing early promise is AI assistance at the pattern recognition level — identifying features like nuclear-to-cytoplasmic ratio, chromatin pattern, and mitotic figures in a prepared sample. These tasks map well to image classification. The early research is encouraging. It doesn't yet support clinical use without expert review.
Commercial cytology tools: what already exists
AI-assisted cytology isn't purely a future prospect — commercial tools are already in use at some practices. The most established applications are automated blood cell differential counting and fluid analysis. Some platforms are beginning to expand into mass aspirate classification for common tumor types in companion animals.
The scope of what these tools have been validated to do is narrower than their marketing sometimes suggests. Automated counting in a well-prepared blood film is a very different problem from classifying a poorly cellular aspirate from an unusual species. Before relying on any commercial cytology AI tool, it's worth asking specifically what it was validated on, in which species, and against what reference standard.
The oversight principle applies here too: AI cytology tools built without veterinary pathologist input carry the same risk as any AI tool applied without species-specific logic — confident output that doesn't account for what the model was never trained to recognize. AI assistance reviewed by a qualified interpreter is the responsible path. A cytology AI output is a starting point, not a conclusion.
Biopsy site selection: the highest-impact application most people aren't talking about
Where exactly you sample a lesion has an enormous effect on whether the result is useful. Necrotic centers, reactive edges, and areas of secondary inflammation can all produce a result that reads as non-diagnostic — or worse, misleading. An experienced clinician learns over time to read gross morphology and choose accordingly. A less experienced one may not — and no amount of pathology expertise on the other end can recover information that was never in the sample.
AI tools that analyze lesion images and flag likely viable versus non-viable tissue zones are in early development. The concept is sound. The evidence base in veterinary species is still limited, and this remains an area of active investigation rather than a current clinical standard.
The submission decision: AI as a prompt to act
One of the most underappreciated things AI can do at this stage is simply prompt the submission decision in the first place. A meaningful proportion of actionable lesions in veterinary practice are never sampled — not from lack of skill, but because the clinical picture didn't quite reach the threshold that triggers submission.
A tool that generates a morphology-informed differential list at the point of gross evaluation makes the stakes of the decision visible. When a tool surfaces malignant possibilities for a lesion a clinician might otherwise have elected to monitor, it creates a documented, transparent prompt to consider submission. That shift — from intuition to recorded reasoning — has real implications for continuity of care and medical record quality.
What AI cannot fix: Sample quality is still a function of technique, fixation, and handling — none of which AI currently influences. A well-reasoned submission decision followed by a poorly prepared sample still produces a non-diagnostic result. AI tools work best alongside strong foundational practice in sample collection and submission.
Where the field is moving
The direction is toward integration — tools that carry clinical context forward from the exam room to the laboratory rather than letting it disappear in between. The technical infrastructure for this is beginning to exist. The species-specific, validated content that should power it is still being built.
What clinicians can reasonably expect from AI in the near term is assistance — with pattern recognition, differential generation, and prompting submissions that might otherwise not happen. What they shouldn't expect is a substitute for the interpretive expertise that has always been at the center of veterinary diagnostics.
Next in this series: Computational pathology — what AI sees under the microscope, where it's already performing, and what it still gets wrong.
Written by Eric Snook, DVM, PhD, DACVP · Vetopathy LLC