The critical role of standards for AI in digital pathology

Abstract

Background

The field of pathology has not yet fully realized the potential of artificial intelligence (AI) and digital pathology. Adoption must be driven by demonstrable utility, and successful implementation depends on interoperability and sustainability, which require established standards. We address the imminent challenges facing the field of AI in digital pathology, which currently suffers from a lack of coordinated and adopted standards.

Methods

We conducted several roundtable discussions with key opinion leaders from multiple sectors across the healthcare ecosystem. Based on how standards are used, we distinguish different areas of practice (relevance, endorsement, and utility) and emphasize the importance of standards.

Results

Our roundtable discussion centered on one key theme: successfully implementing AI in digital pathology depends on achieving a certain level of uniformity across practices. We derive an approach to describe the critical role of standards consisting of seven interdependent areas of practice: value recognition, existing standards, dependencies for AI, failures, management of standards, trends, and a roadmap for accelerated and sustainable adoption. The promise of standards and our approach can be understood as the interconnection of these areas. We address imminent challenges surrounding the field of digital pathology by providing an approach for interoperable, coordinated, and sustainable use of standards across diverse practice settings.

Conclusion

With the concepts and frameworks outlined in this article, we highlight the importance of standards in pathology and their crucial role in driving computational advances and enabling AI solutions to enhance patient care.

Introduction

Standards are essential building blocks of modern healthcare.^1,2 In lab medicine, a standard is typically characterized by: (i) a consensus-driven specification, (ii) an established governance or maintenance process, and (iii) clear expectations for conformance or implementation. In pathology, standards enable consistency in slide size, staining protocols,³ magnification levels, and image formats.4, 5, 6, 7 These conventions are so embedded in daily workflows that they often go unnoticed.⁸ Yet, despite this reliance, the configuration and implementation of standards remain highly variable across institutions and geographies, shaped more by local precedent than by shared frameworks.9, 10 This variability poses increasing challenges as the field undergoes a digital transformation11, 12, 13 (Fig. 1).

Fig. 1.

Pathologists use many “Standards”. Standards in pathology fall into three large groups: structural (e.g. setting or size), process (e.g., standardized activities, standard operating procedures), or outcomes (standardized evaluation of an activity). Each group is guided by foundational, governance, and interoperability standards. However, here we focus on standards directly related to realizing AI in digital pathology.

The rationale for broad adoption is not technological inevitability, but rather the longstanding principle that standardized and reproducible practices (central to both clinical practice, clinical trials, and diagnostic pathology) are what allow innovations such as AI-derived biomarkers to deliver measurable utility for patients and society.14, 15, 16, 17

Given the proliferation of annotation formats across the field (e.g., CSV, XML, JSON, or GeoJSON), establishing clear, harmonized guidelines for the encoding, exchange, and provenance of annotations is essential to enable interoperability, reproducibility, and scalable AI development in digital pathology.

Standards are best adopted in incremental, value-aligned steps, not as a single comprehensive package. However, not all parameters require domain-specific standards; where higher-precedence standards such as SI units already provide an adequate, continuous representation of a variable, additional standardization may be unnecessary. In practice, a lab director can distinguish a true standard from a “standard-adjacent” tool by assessing whether it reflects formal consensus, includes explicit conformance requirements, and is maintained by a recognized standards body; benchmarking utilities such as MedHELM or HistoQC support evaluation and quality monitoring but do not constitute validation standards themselves.

Digital pathology and the integration of artificial intelligence (AI) have introduced new complexities to the discipline¹⁸; complexities that traditional standards are not fully equipped to address.11, 12, 13, 19 AI, and in particular machine learning, depends on structured, high-quality data to train, validate, and deploy algorithms that emulate diagnostic decision-making.20, 21 This dependency intensifies the need for interoperable formats, semantic consistency, and quality control across the data lifecycle.22, 23 However, current data collection practices in pathology are fragmented, inconsistent, and largely unstructured, creating a fundamental misalignment between requirements and workflows.²⁴

The transformative potential of AI in pathology is widely acknowledged, yet the underlying infrastructure required to support it remains underdeveloped. Whereas individual standards (e.g., DICOM for imaging,25, 26 ICC for color calibration,27, 28 or IHE/PaLM for data workflows²⁹) have made progress in specific domains,23, 30, 31, 32 there is no comprehensive, up-to-date resource that maps out the standards ecosystem in computational pathology, particularly with respect to AI implementation.¹

We considered this absence a critical gap. Unlike in radiology or genomics, where standards have been widely adopted and are actively maintained, pathology lacks a unified reference framework that links standards development to AI-readiness, regulatory alignment, and clinical implementation. Without such a resource, institutions, vendors, and regulators may struggle to coordinate efforts, assess compliance, or scale solutions across systems.

The aim of this work is to address this gap by providing a structured overview of standards in computational pathology, with a focus on their relevance for AI. We examine the types, functions, and current state of applicable standards, highlight critical areas for development, and outline how standards can serve as enablers (i.e., not barriers) to safe, scalable, and interoperable AI integration. We aim to support informed decision-making by stakeholders across domains and contribute to a shared foundation for innovation in digital pathology and AI in research and patient care.

Methods

This work employed a structured narrative review approach to characterize existing and emerging standards relevant to computational pathology, with an emphasis on their role in enabling AI. The objective was not to evaluate diagnostic performance or conduct a systematic review, but rather to map the standards landscape and highlight design, implementation, and maintenance considerations.

Scope and selection criteria

We focused on standards that are directly or indirectly applicable to digital pathology workflows, image analysis, and AI model development. Inclusion criteria for standards were: (1) relevance to data capture, processing, or interoperability in pathology, (2) endorsement or recognition by regulatory, professional, or standards development organizations, and (3) potential utility for AI-enabled applications.

Data sources and expert input

Sources included peer-reviewed literature, publicly available technical specifications, regulatory guidance documents (e.g., FDA-recognized consensus standards), and outputs from international standards bodies (e.g., ISO, DICOM, ICC, and IHE). Additional insights were provided by domain experts through collaborative contributions from members of professional societies, industry, and regulatory science communities involved in standard-setting or AI implementation.

Organization and analysis

Identified standards were grouped into four categories based on function: (1) foundational (e.g., data structure and image formats), (2) governance (e.g., quality management and validation protocols), (3) interoperability (e.g., communication between systems), and (4) application-specific standards (e.g., AI explainability, performance monitoring). For each, we assessed scope, adoption status, relevance to AI, and known challenges or limitations. The analysis also considered alignment with regulatory frameworks and future trends such as federated learning and synthetic data use.

Results

Existing standards in digital pathology

A range of foundational and interoperability-focused standards already exist within digital pathology. Among the most widely recognized is the Digital Imaging and Communications in Medicine (DICOM) standard, originally developed for radiology and later extended by Working Group 26 to address pathology-specific requirements. These include support for whole-slide imaging (WSI), image metadata, and hierarchical image tiling. DICOM enables standardized image exchange across systems and has become increasingly important with the growth of telepathology and AI validation studies that rely on multicenter image sharing. DICOM leverages SNOMED CT as a standardized ontology of medical concepts to support semantic interoperability by enabling consistent annotation and tagging of clinical content.

The International Color Consortium (ICC) standard plays a key role in ensuring consistent color representation across devices, which is critical for reproducibility in digital image analysis.33, 34 ICC profiles are used to align the color output from scanners and displays, thereby reducing variability in computational image interpretation, particularly in tasks such as PD-L1 or HER2 scoring or tumor segmentation.35, 36

The Integrating the Healthcare Enterprise – Pathology and Laboratory Medicine (IHE PaLM) initiative focuses on workflow and data integration, leveraging existing standards whenever possible to solve interoperability challenges. The efforts of this initiative established technical frameworks that enable interoperability between laboratory information systems (LISs), electronic health records (EHRs), WSI scanners, digital pathology viewers, image archives, and image analysis software.³⁷ These integration profiles support standardized communication protocols and play a pivotal role in scalable deployment of AI tools.

An overview of relevant standards in digital pathology is provided in Table 1. While we cannot cover all standards, one key result is the recognition of the value in applying existing standards rather than re-inventing them.

Table 1.

Overview of standards in digital pathology (selected).

Standard	Description	Application	Example use case	Organization	Category
DICOM	DICOM is the international standard to transmit, store, retrieve, print, process, and display medical imaging information. The DICOM Working Group 26 supports and develops the standards for the pathology domain.	Storage and communication of digital pathology images	SP146; slide exchange in multisite AI model validation	DICOM (WG-26)	Interoperability
OME-TIFF	OME-TIFF is an open standard file format developed for storing, sharing, and analyzing biological image data, including multidimensional image formats.	Image data structuring and sharing	Multichannel microscopy data in AI training	OME Consortium	Foundational
ICC profile	Defines color characteristics of input and output devices and ensures consistent image appearance across devices and institutions.	Color normalization and display consistency	Standardized PD-L1 image review across platforms	ICC	Foundational
JPEG	Standardized image compression format widely used in medical imaging systems for efficient storage and transfer.	Image compression and exchange	Efficient WSI storage with minimal loss	JPEG (ISO/IEC)	Foundational
CPT	CPT codes are used to report medical, surgical, and diagnostic procedures and services.	Billing and reimbursement	AI-driven billing code generation for digital reads	AMA	Governance
ICD, SNOMED	ICD and SNOMED CT are used for standardized disease classification and clinical terminology.	Clinical terminology and disease classification	Standardized annotation of training datasets	WHO/IHTSDO	Application-specific
Antibody classification	Defines categories and nomenclature for antibodies used in diagnostics and therapeutics (e.g., PD-L1 classification standards).	Biomarker identification and categorization	PD-L1 scoring in immunotherapy response prediction	WHO/commercial panels	Application-specific
AJCC/TNM	AJCC/TNM is a global standard for cancer staging based on tumor size, lymph node involvement, and metastasis.	Cancer staging and treatment planning	TNM classification in prognostic AI tools	AJCC	Application-specific
HGNC	HGNC provides unique symbols and names for human gene nomenclature, including variants relevant to molecular pathology.	Genomic variant reporting	Gene panel output mapping in variant databases	HGNC	Application-specific
Staining	Standard protocols for hematoxylin and eosin staining, ensuring reproducibility and diagnostic consistency.	Staining consistency	H&E-based AI model for cancer detection	CAP/pathology consensus	Foundational
Processing	FFPE processing standardizes tissue preparation for microscopy and molecular analysis.	Tissue preparation	FFPE-based image archive for algorithm validation	ASCO/pathology guidelines	Foundational
Magnification	Magnification standards (e.g., 20× vs. 200×) ensure consistent interpretation and computational analysis of image data.	Image interpretation consistency	Scaling ML algorithms across magnifications	CAP/DICOM	Foundational

Abbreviations: DICOM: Digital Imaging and Communications in Medicine; OME: Open Microscopy Environment; ICC: International Color Consortium; CPT: Current Procedural Terminology; AMA: American Medical Association; ICD: International Classification of Diseases; SNOMED: Systematized Nomenclature of Medicine; WHO/IHTSDO: World Health Organization/International Health Terminology Standards Development Organization; AJCC: American Joint Committee on Cancer; HGNC: HUGO Gene Nomenclature Committee; CAP: College of American Pathologists; FFPE: formalin-fixed, paraffin-embedded; ASCO: American Society of Clinical Oncology.

AI-specific requirements

For AI to be successfully implemented in pathology, standardization is not optional; it is foundational. AI model training requires large volumes of consistent, well-annotated data. Variability in staining protocols, tissue processing, and scanner configurations can introduce confounders that limit model generalizability. Algorithms trained on data from a single institution may perform poorly when applied to external datasets, underscoring the need for harmonized pre-analytical and analytical procedures. In other words, we identified a certain degree of dependency for AI to avoid failures.

WSIs are sensitive to artifacts such as blur, tissue folds, or air bubbles, which can distort model performance. Several groups have proposed quality control standards for WSIs to address these issues,³⁸ enabling cleaner training datasets and more reliable model outputs. Standardized formats such as OME-TIFF and DICOM facilitate structured data capture and metadata embedding, which is essential for version control and reproducibility in AI workflows. Furthermore, interoperability standards allow for the aggregation and harmonization of data from multiple sites. This is a prerequisite for federated learning and multi-institutional validation studies. Without standardization, the lack of cross-site generalizability remains a persistent barrier to clinical deployment.

Regulatory and organizational alignment

The regulatory ecosystem is increasingly acknowledging the importance of standards in the deployment of AI tools in pathology. The U.S. Food and Drug Administration (FDA), through its Digital Health Center of Excellence, has issued guidance emphasizing the role of standards in evaluating software-based medical devices, particularly those involving continuous learning or real-world data.

The Digital Pathology Association (DPA) also plays a pivotal role in advancing the standardization and regulatory clarity of digital pathology tools.³⁹ Through its Regulatory and Standards Task Force, the DPA actively engages with stakeholders (including industry, academia, and regulatory agencies) to promote interoperability, clarify approval pathways, and support the adoption of standards that facilitate safe and effective use of AI in clinical practice (https://digitalpathologyassociation.org/digital-pathology-association-regulatory-committee; last accessed on 07/02/25).

The College of American Pathologists (CAP) has developed digital pathology checklists and validation protocols that cover aspects such as pre-analytic image handling, image analysis validation, and quality control.10, 40 These41, 42 resources are especially relevant for labs implementing AI-assisted diagnostics and seeking regulatory approval or accreditation. The CAP also convenes both the DICOM working group 26 and the IHE/PaLM initiative.

Several organizations are directly involved in standard development and ecosystem coordination. The Pathology Innovation Collaborative Community (Plcc), formerly known as the Alliance for Digital Pathology, operates in collaboration with the FDA and other stakeholders to advance regulatory science through shared standards.⁴³ Additional contributors include the Medical Device Innovation Consortium (MDIC) and the CAP Council on Informatics and Pathology Innovation (CIPI), both of which support the development, dissemination, and adoption of computational pathology standards. These organizations collectively emphasize that standards require management through sustainable organizations.

The duality of standards

This landscape reflects a dichotomy in how solutions are developed and adopted: the customized approach versus the standards-based approach (Table 2). Customized approaches offer flexibility, speed, and alignment with local workflows, often favoring rapid innovation and iterative development. However, they tend to be difficult to scale and sustain over time. In contrast, standards-based approaches require greater upfront investment, coordination, and domain expertise, but yield long-term benefits through interoperability, auditability, and ecosystem compatibility. Whereas the former may appear more agile, the latter provides the structural foundation necessary for scalable AI integration across diverse clinical settings, again emphasizing the needs of developing and managing standards.

Table 2.

The dual nature of working with standards.

Customized approach	Standard-based approach
Highly individualized to local setting and question	Generalized solution that applies to other settings and questions
Short-term solution	Long-term solution
Example: .svs files and .xls	Example: DICOM
Learning while doing	Learning then doing
Heavy emphasis on functional relationships of components	Heavy emphasis on technical specifications of the components and their interaction
Initial low cost and fast adoption	Larger initial effort
Fast changes and iterative development	Changes of the standard take time (e.g., workgroups, backwards compatibility, and vocabulary)
Rarely sustainable over a long-time frame	Sustainable and large set of tools
Commitment to an approach for a solution	Commitment to “a” standard requires domain knowledge
Non-standardized solutions appear more innovative at first but cause long-term problems	Standards appear restrictive and constraining at first but result in long-term pay-off
Emphasis on quick adoption	Emphasis on interoperability
Organization: local control	Organization: committee/nominations
Strengths: flexibility, simplicity, and low-cost	Strengths: compliance, auditability, and compatibility

Emerging needs and future trends

As AI applications mature, new challenges are emerging that demand dynamic and forward-compatible standards. The integration of real-world data into AI model development and regulatory evaluation introduces a need for consistent data structures, semantic frameworks, and performance monitoring systems. Tools that adapt over time—such as those based on continuous learning—require formal mechanisms for post-market surveillance, auditability, and explainability. Federated learning, in which models are trained across decentralized data sources without sharing raw data, offers an attractive path forward for collaborative AI development. However, this paradigm depends heavily on common data formats, aligned ontologies, and standardized performance metrics across institutions. The use of synthetic data to augment real-world datasets is another area of rapid evolution. Trials such as the VICTRE study in radiology have demonstrated the feasibility and value of synthetic cohorts for regulatory evaluation. In pathology, the generation, labeling, and validation of synthetic images will similarly require standard frameworks to ensure utility and comparability. Lastly, the growing emphasis on explainable AI (i.e., where algorithmic outputs are interpretable and traceable to input data) will necessitate standards for output formatting, provenance tracking, and model transparency. These features are essential not only for regulatory approval but also for clinician and patient trust. Following trends and developing a roadmap for adoption is key.

To illustrate the multifaceted role of standards in these pathology innovations, we developed a conceptual overview of the foundational dimensions of effective standards; it is not a stepwise adoption roadmap (Fig. 2). In practice, labs adopt standards iteratively, prioritizing those with the highest regulatory, operational, or interoperability value while remaining adaptable as standards evolve.

Fig. 2.

Core functional dimensions of standards as a concept. Standards in digital pathology and AI span technical, regulatory, operational, and ethical dimensions. This figure illustrates 11 foundational principles that guide the development, implementation, and sustainability of standards. From governance and regulatory alignment to harmonization, versioning, and equity, the diamond-shaped pathway highlights how standards must be grounded in existing frameworks, evolve through experience, and offer a clear value proposition. Effective standards require both top-down oversight and bottom-up practicality to ensure they are adopted, maintained, and trusted across diverse clinical and technological ecosystems.

Discussion

Here, we provide a structured synthesis of existing standards relevant to digital pathology and their role in enabling the adoption and scalability of AI. By categorizing standards into foundational, interoperability, governance, and application-specific domains, we illustrate how standards underpin every stage of AI deployment, from data acquisition and model training to regulatory validation and real-world implementation. We derive an approach to describe the critical role of standards consisting of seven interdependent areas of practice: value recognition, existing standards, dependencies for AI, failures, management of standards, trends, and a roadmap for accelerated and sustainable adoption. This overview highlights both the technical and organizational scaffolding needed to translate digital pathology innovation into sustainable clinical practice.

Our work fills a notable gap: whereas various standards have been proposed or adopted across domains, no single resource has consolidated and analyzed their relevance specifically for computational pathology and AI. The novelty of this work lies in its integrative scope. Prior reviews have focused on individual components (e.g., imaging formats, data exchange, quality control, or color calibration); however, to our knowledge, none have examined the broader system of interdependent standards necessary for scalable, interoperable AI. Furthermore, this overview is grounded not only in technical documentation but also in regulatory frameworks, real-world deployment scenarios, and collaborative initiatives like DPA, Plcc, MDIC, and CAP-CIPI. By doing so, we bridge the conceptual divide between technical specifications and clinical implementation.

There are limitations to this work. As a narrative review, our findings are subject to selection bias and may not capture every emerging or regional standard. The field is rapidly evolving, and certain initiatives (e.g., in synthetic data governance or explainable AI) are still under development. Additionally, standards are often embedded in institutional workflows or vendor systems in undocumented ways, making comprehensive coverage difficult. We also acknowledge that the utility of a given standard can vary by region, resource level, and clinical application. We consider these limitations acceptable when considering the main aim: providing a concise and meaningful starting point.

Looking ahead, the future of AI in pathology will increasingly depend on dynamic, multipurpose standards capable of supporting both innovation and regulation. As AI models evolve from locked to continuously learning systems, there will be a growing demand for standards that enable real-time auditing, version control, and model explainability. Federated learning and synthetic data offer tremendous promise but will require harmonized data structures and robust ontologies. Importantly, these standards must not only be technically sound but also accessible to developers, regulators, and clinicians alike. The convergence of imaging, genomics, and spatial biology will further accelerate the need for cross-domain standards that support integrated diagnostics.

In summary, the development, implementation, and maintenance of standards are critical to realizing the promise of AI in digital pathology. Whereas the creation and adoption of standards may appear restrictive at first, they serve as long-term enablers of interoperability, scalability, and trust. We hope our work serves as a foundational reference for multiple stakeholders across industry, academia, and regulation; encouraging informed participation in the standardization process and aligning innovation with sustainable positive impact on patient care.

Funding

Supported by the Digital Pathology Association.

Declaration of competing interest

L. P. is on the medical advisory board for Ibex and NTP, consults for Hamamatsu and AiXMed, and is a cofounder of LeanAP and Placenta AI. M. Herrmann is a Employee and shareholder of Roche. All other authors have no conflicts of interest to disclose.