Artificial intelligence in pathology today helps rather than replaces. A pathologist and Jaroslav Vohánka spoke about experiences from Krajskej nemocnice Liberec and Roche’s approach to developing and deploying algorithms. They summarized the benefits for diagnostics as well as the obstacles – from error rates and legislation to the integration of hospital data.
Safety, error rates, and the role of the pathologist
The pathologist always has the final say, so overall error rates hinge on their decision. The aim of the algorithms is to reduce routine mistakes – for example, to flag a suspicious lesion that a person might overlook when distracted. Studies report that the average error rate of human assessment hovers around 3–5 %, while algorithms come in below that threshold. AI thus adds another layer of control, not a replacement for the expert.
There is also a clear benefit in harmonizing assessments, for example in Gleason grading of prostate cancer. The algorithm evaluates consistently regardless of whether it's "Monday afternoon" or "Wednesday morning," and serves as a second look when the pathologist is unsure. In effect, it "lends" experience to younger colleagues, making learning easier and speeding up the work. In a digitized workflow, this helps handle a growing number of samples without having to increase staffing.
Barriers to implementation and what data integration will bring
The biggest brakes are legal and organizational: GDPR, the difference between anonymization and pseudonymization, or hosting servers outside the EU. Contractual and approval processes for linking to third-party services (e.g., cloud) take months to years; in the case mentioned, implementation took roughly a year and a half. Even though the servers for Roche solutions are located in the EU, involving additional providers brings complex negotiations about data protection and jurisdiction. The good news is that feasible and safe configurations exist and are already used in practice.
Alongside technology, integration is key: to connect digital pathology, genetics, and oncology and create a secure environment for joint decision-making. The goal is to give the oncologist a tool that, based on pathology, NGS, and clinical data, proposes scenarios for personalized treatment and an overview of recommendations from studies. The automatic "patient summary" also looks promising – AI that condenses extensive documentation into a concise overview with the most important data. Looking ahead, the plan is to connect to national platforms so that the patient doesn't travel with paper documentation, but with securely shared data.
