What is actually holding back AI adoption in pathology laboratories?
In a detailed discussion with Professor Hamid Tizhoosh, Professor of Biomedical Informatics at Mayo Clinic, the central question is the gap between inflated expectations and clinical reality. Foundation models and large language models are still far from meeting five non-negotiable requirements: accuracy, consistency, speed, efficiency, and robustness.
The problem is not a lack of research or funding. The problem is that much of what appears in conference talks and press releases is not tested under real clinical conditions. Models that reach 95% in laboratory tests may fail when faced with the real variation in tissue, staining, and the uneven quality of digital slides.
The five requirements: why does AI fail to meet them?
Accuracy
A model that is correct in 90% of cases may be acceptable for recommendation or triage, but it is not acceptable for final diagnosis. A 10% error rate means that one patient in ten receives the wrong diagnosis. In pathology, that number is catastrophic. The accuracy required in clinical practice is close to 99%, a level that no foundation model has yet approached in independent evaluations.
Consistency
The same digital slide may receive a different diagnosis each time the model processes it. This problem is worse than inaccuracy because it is unpredictable. A pathologist may disagree with a colleague, but that disagreement usually follows an understandable pattern. Model variability is random and cannot be anticipated. A recent study showed that some models give different results for the same WSI when the image is rotated by 90 degrees.
Speed
Real-time processing is not a luxury. A laboratory that processes 300 slides a day needs results within minutes, not hours. Many research models require computing infrastructure that makes real-time use practically impossible. The gap between inference speed in a research setting and in production is large.
Efficiency
A model that needs 16 GPUs to analyze one slide is not a realistic solution for most laboratories. Efficiency means obtaining the required results with the fewest possible resources. That requires an engineering approach very different from what is often built in academic laboratories, where compute resources may be treated as unlimited.
Robustness
A model trained on data from one scanner may fail completely with a different scanner. Variation in staining protocols, preparation quality, and tissue thickness makes robustness a central problem. Every laboratory knows that a small difference in H&E staining can produce a large difference in model output.
The multimodal approach: is it the solution?
Tizhoosh discusses the possibility that a multimodal approach could close some of these gaps. Instead of relying on the image alone, this approach combines histology data with clinical, genomic, and immunologic information. The idea is not new, but applying it to large foundation models is what makes it clinically relevant.
The challenge is how to represent all these modalities consistently. IHC data are not the same as WSI data, and an immune response cannot be reduced easily to a numerical vector. A model that handles multiple modalities well needs training on carefully labeled data from each modality. That requires collaboration between institutions that usually do not share their data.
What should pathologists demand?
The responsibility does not fall on developers alone. Pathologists are the end users, and they have the right to ask hard questions before adopting any technology.
Ask about the training data: where did it come from? How many laboratories contributed? Does it include ethnic, age, and geographic diversity? Data from one hospital in North America do not represent the world.
Ask about independent evaluation. A study published by a developer about its own product is not enough evidence. Look for evaluations by third parties and by laboratories that used the model under real conditions.
And ask about the review mechanism. What happens when the model is wrong? Is there a system for recording and analyzing deviations? Or is the model a black box that cannot be audited?
A technology that cannot be audited should not be adopted. The rule is simple. The pathologist is responsible to the patient, and that responsibility cannot be handed over to a tool whose decisions cannot be understood or reviewed.