How close are we to a point where AI systems can generate clinically actionable results without a pathologist reviewing every case?
The current trajectory of computational pathology does not aim to replace the pathologist or create systems where pathologists do not review cases. Instead, the field is moving toward a collaborative system, where the pathologist’s judgment is still required, but there is more focus on efficient case management and complex diagnostic integration.
Rather than using them in diagnostics, AI algorithms are best placed to measure biomarker expression, providing more precise quantification of biomarkers than humans.
What technical or regulatory hurdles remain for AI diagnostics?
The shift toward integration with AI-assisted tools faces several regulatory hurdles. Regulators, such as the FDA, require medical device manufacturers to define these tools as parts of an "end-to-end" system that includes the entire workflow – from tissue staining and scanning to the final algorithm-derived result. Implementation of a complete system for an AI algorithm may represent an operational burden to the lab.
Technical burdens include the need for heightened pre-analytic control, such as various fixatives or fixation times, standardized staining protocols, and section thickness, all of which may impact algorithm output. System hurdles include complexity, infrastructure, and cybersecurity. Implementation barriers center around cognitive and educational shifts for pathologists and oncologists, and workflow adoption.
The adoption of AI as part of clinical testing will likely become more mainstream as digital pathology is more widely implemented, and as the pathology field gains more experience and evidence on available AI tools.
How effectively can current AI models evaluate and score cancer biomarkers compared with expert pathologists?
AI models allow pathologists to move from what can be subjective visual estimation of biomarker expression in tumors, to more precise quantification. Pathologists currently use qualitative or semi-quantitative scoring, which involves visually categorizing results into "bins," such as 0, 1+, 2+, and 3+.
Manual scoring by pathologists is inherently susceptible to both interobserver variability (different pathologists assessing the same slide differently), and intraobserver variability (the same pathologist scoring a slide differently at different times). AI models aim to replace this scoring method with quantitative scoring, some models with the ability to count every tumor cell and pixel, to produce exact continuous metrics.
AI models may better quantify biomarker expression in tumor cells, minimizing the subjectivity inherent in human assessment. A particular struggle is differentiating immunohistochemistry (IHC) expression in predictive biomarkers, especially in borderline cases. For example, with a 50 percent IHC scoring cutoff, distinguishing the range of expression near a threshold – say 45 to 55 percent – is challenging, particularly given these decisions are important in selecting treatment. In addition, manual assessment can lead to fatigue when examining large sections from tumor excisions.
...and where do such models still fall short?
While AI may offer greater precision, it does not function autonomously and faces operational and cognitive barriers, including its dependency on upstream quality. Use of these tools requires a cognitive shift in mindset for pathologists and oncologists, as well as acquiring AI literacy and training, and infrastructure updates for laboratories.
Another key point is that we currently accept manual pathologist assessment as the “ground truth” for scoring, with AI largely viewed as an adjunctive aid. With the recent addition of vision language and foundation models to pathology AI, model performance is expected to increase, and variations between models will likely be reduced. These improvements will eventually converge into an AI-generated ground truth that will be able to be trusted by pathologists.
Could AI tools ultimately establish a new standard for reproducibility in companion diagnostics?
Yes. The diagnostics industry is moving toward this new standard, where results are based on more objectively quantified measurements wherever possible, allowing pathologists to spend more time focusing on complex interpretation of the holistic case.
Reproducibility of biomarker results with AI tools can be maintained by validating and monitoring the entire workflow – including tissue staining, scanning, and the algorithm itself – ensuring that all components perform within defined parameters.
By incorporating future technologies, such as the VENTANA TROP2 (EPR20043) RxDx Device (product in development), pathology can employ a standardized quantitative language for biomarker expression that is not reliant on traditional visual estimation. In other words, these AI tools are capable of identifying clinically relevant protein expression levels that can predict therapeutic response, which traditional visual scoring cannot see.
While this AI technology provides biomarker quantification to improve reproducibility, pathologists remain in the loop to ensure appropriate quality control and sufficiency in multiple steps. This collaborative model for computational pathology devices maintains the pathologist's central role while utilizing AI to enforce high-precision standards.
If an AI system independently generates a biomarker result that determines whether a patient receives a targeted therapy, who ultimately bears responsibility for that decision?
Pathologists and laboratory staff oversee and control the quality of the inputs for AI, including pre-scan slide preparation, post-scan image inspection, and QC analysis, which may be automated, and typically includes artifact detection and evaluation of color consistency among others. Images scanned on different scanners, staining done on different staining platforms, and artifacts, may all impact AI results, highlighting the great importance of QC when employing AI tools.
AI tools in diagnostic pathology are not currently considered to be autonomous decision-makers, but rather advanced precise biomarker quantifiers, designed to evaluate and score protein expression in tissue samples to assist the pathologist. The pathologist's role is not replaced, instead, it is reframed as collaborative, with the pathologist retaining control through several key checkpoints.
Ultimately, the pathologist is responsible for complex diagnostic integration, including applying professional judgement for inputs and outputs of the algorithm, and incorporating this within the context of the patient’s specific clinical history — a crucial context that AI algorithms cannot consider. Accordingly, the pathologist has the final authority to accept or reject the algorithm results. If the pathologist disagrees with the generated biomarker status, they are empowered to reject the analysis, and proceed with other options.
As AI becomes more integrated into biomarker assessment workflows, how is the role of the pathologist changing today?
Pathologists may be required to shift from qualitative visual estimation to AI assisted quantitative scoring methods. Conventional companion diagnostics typically rely on ordinal scoring systems that can fail to capture the spectrum of protein expression or to identify subtle sub-cellular localization. This more precise analysis may require quantitative metrics, such as the normalized membrane ratio being investigated in future technologies.
Pathologists continue to provide essential, specialized expertise to the process, such as ensuring the specific algorithm is being applied appropriately to proper subtypes of tumor. The pathologist is responsible for ensuring the inputs to the computational system are diagnostically relevant, including verifying adequate tumor, adequate staining, and artifact-free scanning.
The pathologist monitors how the algorithm performs and has the power to exclude specific areas of the tissue that the algorithm might have incorrectly selected for analysis. After the algorithm generates a biomarker status, the pathologist can accept or reject the result, and, if indicated, seek a second opinion.
How important is explainability? Do pathologists need to understand exactly how an algorithm reached a particular biomarker score, or is demonstrated clinical performance sufficient?
Pathologists don't need to understand every aspect of the technology powering a microscope, and by extension they don't need to be able to explain every aspect of the technology of an algorithm. However, they need to have some understanding of how the algorithm is trained and validated, and how it arrives at a result. Pathologists want to see objective evidence and demonstrated clinical performance, which contributes to the overall explainability of the algorithm.
Could future AI-driven diagnostics move beyond companion diagnostics altogether, predicting treatment response, prognosis, or molecular alterations directly from histology slides without requiring traditional biomarker testing?
In the near future, we’ll see a shift toward computational companion diagnostics, representing an evolution of the current model of manually assessed companion diagnostics. In the longer term, novel AI technologies are emerging with the potential to predict biomarker expression and therapy response based upon computational analysis of the H&E alone. A study in the Journal of Clinical Oncology highlights how deep learning models exploit the rich morphology of H&E slides to predict rare genetic mutations in lung adenocarcinoma.
Computational pathology companion diagnostic assays will rely on an integrated workflow where the IHC staining and AI tools are both essential components to the final algorithm-derived result. At present, IHC continues to serve as the cornerstone of the companion diagnostic framework.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
