I probably saw this post via someone at Linked-In, but I've mislaid how I heard about it.
Worth reading, a new open-access paper, which projects into the future and asks if pathologists will be "partners" or "bystanders" to AI.
Find the September 2025 paper by El-Khoury and Zaatari online at Diagnostics, here:
https://www.mdpi.com/2075-4418/15/18/2308
In addition to the main story, the article offers an excellent bibliography of nearly 100 references, some of them hard to find (history of microscopes), many others up-to-date through 2025.
AI CORNER
###
El-Khoury and Zaatari provide a sweeping, historically anchored analysis of how digital pathology and AI are jointly reshaping diagnostic practice, culminating in the provocative question of whether pathologists may eventually become “bystanders” in workflows they once wholly governed. They trace the discipline’s evolution from early microscopy and microtechnique innovations to contemporary whole-slide imaging (WSI)—establishing digital pathology as the essential substrate upon which AI systems now depend. This historical framing underscores that AI represents not merely another tool but a structural inflection analogous to the rise of cellular pathology itself.
The authors categorize current AI systems into task-specific deep learning models and emerging general-purpose “foundation models”, noting that the former already deliver expert-level or superhuman performance in constrained tasks—metastasis detection, Gleason grading, MSI inference, mutation prediction, biomarker quantification—supported by regulatory milestones such as Paige Prostate Detect (FDA-cleared) and Ibex Galen Prostate (510(k)). These systems reduce interobserver variability, shorten turnaround time, and uncover otherwise-missed cancers. However, they remain narrow, brittle, and dependent on pathologist-labeled data.
Foundation models—e.g., Virchow, GigaPath, PLIP, and the multimodal copilot PathChat—reflect a deeper conceptual shift. Trained with self-supervision on diverse histologic corpora, they operate across tasks, integrate language, and can propose differentials, ancillary stains, and draft reports. Their flexibility suggests that future diagnostic pipelines may not be built around individual algorithms but around generalist, multimodal AI agents capable of interacting directly with clinicians.
A central contribution of the paper is its scenario analysis of a hypothetical pathologist-free workflow, in which AI interacts directly with the referring physician. The authors map each step—from grossing and scanning to AI-generated interpretations—arguing that this model is technologically imaginable but clinically, ethically, and legally fraught. They emphasize key obstacles: domain shift vulnerability, demographic bias, the opacity of deep networks (“black box” problem), and unresolved questions of liability when AI errors cause harm. The risk of deskilling the pathology workforce and undermining quality in rare or ambiguous cases is highlighted as a structural hazard.
Ultimately, the authors outline three trajectories: a symbiotic model (AI augments pathologists), a transformational model (pathologists become diagnostic integrators and AI supervisors), and a disruptive model (AI performs routine diagnosis autonomously). They argue that the profession’s future relevance will depend on active engagement with AI tools, evolution of training programs, and leadership in ethical and regulatory frameworks. The paper concludes that while full replacement is unlikely in the near term, long-term disruption—particularly with domain-specific AGI—cannot be dismissed.
