High diagnostic accuracy alone is insufficient to demonstrate the clinical value of artificial intelligence (AI) in digital pathology, according to an editorial published in the Journal of Pathology Informatics.
Although AI systems have achieved expert-level performance in tasks such as tumor detection, grading, and biomarker assessment, their adoption in routine clinical practice remains inconsistent. The authors argue that successful AI implementation depends on more than algorithm performance, requiring evidence that tools improve diagnostic decisions and ultimately benefits patients.
To address this gap, the authors propose a framework linking six stages of the diagnostic pathway: tissue preparation, image acquisition, AI model performance, laboratory workflow, clinical decision-making, and patient outcomes.
Each stage introduces potential sources of error that may affect subsequent decisions. For example, differences in tissue fixation, staining, and section thickness can influence image quality, while variations between scanners and laboratory procedures may affect AI performance.
Consequently, an algorithm validated in one laboratory may not deliver equivalent results elsewhere. The authors emphasize the importance of external validation using representative patient populations, alongside local testing and ongoing performance monitoring.
Importantly, technical accuracy does not necessarily translate into better clinical judgment. The editorial cites research involving 28 pathology experts in which incorrect AI recommendations overturned previously correct assessments in approximately 7 percent of consultations. This highlights the potential for AI to introduce diagnostic errors through inappropriate reliance on automated advice.
The proposed framework therefore calls for evidence that AI systems integrate safely into laboratory workflows, support appropriate diagnostic decisions, and produce measurable clinical benefits.
For laboratories considering AI adoption, this means evaluating more than published accuracy figures. Assessment should include compatibility with existing systems, performance across different scanners and patient populations, and how pathologists respond to AI-generated recommendations.
The authors also recommend continued monitoring after implementation, particularly when laboratory procedures, software, or patient populations change.
The framework is conceptual and has not yet been validated in prospective studies. Nevertheless, it offers laboratories a structured approach to identifying evidence gaps and assessing whether promising AI technologies are ready to deliver meaningful improvements in diagnostic practice.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
