Why do errors arise in artificial intelligence diagnostic tools in histopathology and how can we minimize them?

Abstract

Artificial intelligence (AI)‐based diagnostic tools can offer numerous benefits to the field of histopathology, including improved diagnostic accuracy, efficiency and productivity. As a result, such tools are likely to have an increasing role in routine practice. However, all AI tools are prone to errors, and these AI‐associated errors have been identified as a major risk in the introduction of AI into healthcare. The errors made by AI tools are different, in terms of both cause and nature, to the errors made by human pathologists. As highlighted by the National Institute for Health and Care Excellence, it is imperative that practising pathologists understand the potential limitations of AI tools, including the errors made. Pathologists are in a unique position to be gatekeepers of AI tool use, maximizing patient benefit while minimizing harm. Furthermore, their pathological knowledge is essential to understanding when, and why, errors have occurred and so to developing safer future algorithms. This paper summarises the literature on errors made by AI diagnostic tools in histopathology. These include erroneous errors, data concerns (data bias, hidden stratification, data imbalances, distributional shift, and lack of generalisability), reinforcement of outdated practices, unsafe failure mode, automation bias, and insensitivity to impact. Methods to reduce errors in both tool design and clinical use are discussed, and the practical roles for pathologists in error minimisation are highlighted. This aims to inform and empower pathologists to move safely through this seismic change in practice and help ensure that novel AI tools are adopted safely.

Early Breast Cancer Risk Assessment: Integrating Histopathology with Artificial Intelligence

“…Identifying order sets and the algorithmic rules within AI has proven to be notably challenging [ 57 , 126 ]. Unexperienced users may develop workarounds that compromise data which leads to jeopardizing data integrity [ 25 , 127 ]. Tissue sample size could also represent a potential issue (i.e., core biopsies), as well as the subjectivity of “hotspot” selection for training the algorithm [ 86 , 127 ].…”

“…Unexperienced users may develop workarounds that compromise data which leads to jeopardizing data integrity [ 25 , 127 ]. Tissue sample size could also represent a potential issue (i.e., core biopsies), as well as the subjectivity of “hotspot” selection for training the algorithm [ 86 , 127 ]. The presence of artifacts stemming from sampling, slide preparation, and slide digitalization can impede computational analysis and lead to erroneous data interpretation.…”

Applications of Large Language Models in Pathology

“…Inadequate understanding of how artificial intelligence models work is a key contributing factor to automation bias [106]. An AI-recommended diagnosis may convince a pathologist to alter a correct diagnosis into an incorrect diagnosis suggested by an AI-based tool [107]. Pathologists need to be informed that AI tools make predictions based on large amounts of mathematical computations, and that these predictions are prone to error.…”

“…Pathologists need to be informed that AI tools make predictions based on large amounts of mathematical computations, and that these predictions are prone to error. Adding a confidence score to predictions may help a pathologist focus on cases where an AI tool is "unsure" of its recommendations [107]. On the other hand, an AI tool may also associate a high confidence score with a wrong diagnosis, so it would be advisable for a pathologist to consult other experts when faced with difficult cases.…”

Closing the loop – the role of pathologists in digital and computational pathology research

“…Strengthening the connection between pathological knowledge generation and the computational field is crucial. In the literature, the lack of communication between pathologists and other scientists can be seen in various examples, including relevant diagnostic classes being missed in the computational training setup (known as hidden stratification), assembly of large case series that fail to represent emerging entity subtypes reinforcing outdated practices, and computational methods not being tested for relevant mimickers and differential diagnoses, to mention a few [ 18 , 19 ]. Pathologists should commit themselves to the AI‐driven transition and regain their central role as tissue experts and evaluators of digital and computational methods.…”