Digital pathology imaging artificial intelligence in cancer research and clinical trials: An NCI workshop report

Abstract

Digital pathology imaging (DPI) is a rapidly advancing field with increasing relevance to cancer diagnosis, research, and clinical trials through large-scale image analysis and artificial intelligence (AI) integration. Despite these advances, regulatory adoption in digital pathology (DP) has lagged; to date, only three AI/ML Software as a Medical Device tool have received FDA clearance, highlighting a validation dataset gap rather than an absence of regulatory pathways. On March 6–7, 2024, the National Cancer Institute held a virtual workshop titled “Digital Pathology Imaging-Artificial Intelligence in Cancer Research and Clinical Trials,” bringing together experts in pathology, radiology, oncology, data science, and regulatory fields to assess current challenges, practical solutions, and future directions. This report summarizes expert opinions on key issues related to the use of DPI in cancer research and clinical trials, including data standardization, de-identification, and the application of Digital Imaging and Communication in Medicine (DICOM) standards. Key topics included data standardization, image quality assurance, validation strategies, AI applications, integration in clinical trials, biobanking, intellectual property, investigators' needs, and lessons from digital cytology and radiology domains. Solutions discussed included adoption of open standards such as DICOM, centralized imaging portals, and scalable cloud-based platforms. The expert consensus outlined in this report is intended to guide the development of DPI infrastructure, standardization, support AI validation, and align regulatory and data-sharing practices to advance precision oncology.

Introduction

Digital pathology imaging (DPI) is increasingly recognized as a fundamental component of cancer research, diagnosis, treatment, and clinical trials. The ability to digitize pathology slides has enabled artificial intelligence (AI)-driven biomarker assessments and discovery, streamlined central pathology reviews, and enhanced translational research.^1,2 In this manuscript, DPI is used as an umbrella term that includes whole-slide imaging (WSI); given that WSI is currently a widely used technology, the terms are sometimes used interchangeably in the literature.

DPI's reach spans clinical diagnostics, translational and molecular pathology, biomarker discovery and validation, drug development, and clinical trial integration, and it intersects with advances in AI-driven image analysis and decision-support tools. Workshop participants emphasized that the optimal deployment of AI in digital pathology (DP) is context-dependent and may occur at multiple points in the workflow. Prospective applications include integration at the scanner or image acquisition stage for real-time quality control (QC; e.g., automated artifact detection) and triage of urgent cases. On dedicated analysis servers, including those running advanced foundation models, AI can detect subtle histological patterns, standardize tumor grading, and quantify biomarkers such as PD-L1 expression. Retrospective use enables large-scale post-scan analyses, such as spatial mapping of tumor heterogeneity, discovery of molecular correlates from WSIs, and validation studies. Finally, integration into institutional image management systems (IMS) and electronic health records (EHRs) allows AI outputs to be linked with genomic, proteomic, and clinical data, ensuring both flexibility and regulatory alignment.³ The Digital Imaging and Communication in Medicine (DICOM) Digital Pathology Working Group has addressed the challenges of standardizing and managing DPI, focusing on hardware, software, slide scanning, de-identification, storage, and sharing of digital files.⁴ Scanners and IMS often use different proprietary file formats (e.g., SVS, NDPI) that may require custom converters or middleware to exchange images. For each combination of scanners (N) and IMS (M), a total of N × M separate bridges are needed, making cross-vendor integration costly to build and difficult to validate and maintain. By having scanners export to DICOM and IMS platforms receive DICOM input via standardized APIs such as DICOMweb™, the number of required bridges drops to N + M. This greatly reduces effort, avoids compatibility issues, and allows images from multiple vendors to flow seamlessly into the same archive, viewer, or analytic pipeline. In order to support legacy analysis applications that do not yet support DICOM but use libraries (such as OpenSlide and BioFormats) that read proprietary formats (like SVS), support for reading (and in some cases, writing) DICOM has already been added to those libraries in common use,^5,6 requiring minimal, if any, change to analysis applications to support tile-based access to DICOM images at multiple resolutions. DICOM WSI standardizes the parameters for WSI storage (tile indexing, pyramid levels, and spatial coordinates), ensuring that algorithms trained on one dataset can be applied reproducibly across others, mitigating incompatibility issues when transitioning from proprietary (e.g., SVS) to DICOM-compliant images. As emphasized in the 2024 NCI Digital Pathology Workshop, this approach both simplifies current workflows and aligns the field with the model of radiology, where DICOM adoption enabled true enterprise-level interoperability.⁷

Complementing these efforts, the Pathology Innovation Collaborative Community (PIcc), through its Alliance for Digital Pathology, seeks to standardize digital pathology processes, promote synergistic regulatory-science efforts specific to DP, and accelerate innovation for patient care.⁸ In parallel, the Friends of Cancer Research (FOCR) Digital PATH initiative convenes pathologists, algorithm developers, regulators, and industry stakeholders to harmonize computational pathology tools, reduce variability in biomarker assessments, and develop reference datasets and validation frameworks.⁹

To harness DPI's full potential in clinical trials and AI-driven research, the workshop convened experts from multiple fields. The overarching goal of the NCI DPI-AI in Cancer Research and Clinical Trials workshop was to assess current challenges, practical solutions, and future directions in pathology. Eight core themes were covered over a 2-day virtual workshop in March 2024¹⁰:

• 1.Digital pathology landscape: Insights from cytology, pathology, and radiology.
• 2.Investigator needs: Essentials for successful DPI implementation.
• 3.AI in research: Applications of AI in DPI and clinical trials.
• 4.Biobank experiences: Standardization efforts and success stories.
• 5.Challenges and solutions: Addressing hardware, software, and data management issues.
• 6.Standardization and quality assurance (QA): Protocols for image acquisition, de-identification, data sharing, and analysis.
• 7.Integration into clinical trials: Effective methods for incorporating DPI and AI.
• 8.Data sharing and IP: Issues around data sharing, intellectual property (IP), and academic and industry partnerships.

To address the above themes, we invited an array of clinical and scientific experts in pathology, radiology, imaging informatics, clinical trials, and IP domains. During the workshop, participants reached consensus on the importance of investing in DPI infrastructure, standardization, and collaborations. Participants collectively emphasized that integrating DPI with unbiased and validated AI has the potential to enhance diagnostic accuracy, patient care, and research outcomes.

To further clarify the distinction between workshop-derived insights and established literature, we provide a summary of major takeaways organized by theme in Table 1.

Table 1.

Workshop-derived insights/recommendations and established literature in digital pathology imaging (DPI)-AI.

Theme	Workshop-derived insights/Recommendations	Supporting literature/References
Infrastructure & workflow	Invest in DPI infrastructure, adopt standardized protocols, and expand IT staffing; optimize pre- and post-digitization workflows; prioritize LIS–IMS integration.	2,11,12
Standards & interoperability	Adopt DICOM (similar to radiology) as the foundation for interoperability Improve metadata exchange between scanners and LIS/EHR Leverage HL7/FHIR integration.	4,13, 14, 15, 16
Image annotation & quantitative pathology	Develop structured annotation frameworks and scalable tools to create AI-ready datasets across institutions.	17, 18, 19
Validation & QA/QC	Calibrate scanners, implement structured QA, and pursue federated validation across sites. Follow CAP guidelines for WSI validation in primary diagnosis (≥60 representative cases, ≥95 % concordance). Expand AI validation to larger, multi-site datasets and integrate uncertainty estimation and out-of-distribution detection.	2,3,11,20,21
Data sharing, privacy & governance	Establish centralized portals (e.g., DPIP) and integrate with NCI repositories (TCIA, IDC) Ensure robust de-identification and address IP/data ownership Expand use of open-source platforms (e.g., DSA); enable federated learning for multi-institutional AI training.	22, 23, 24, 25, 26, 27
Biobanking, clinical trials & AI/Pathomics	Adapt biobanking workflows for biospecimen–image linkage and specimen tracking. Centralize review, QA/QC, and biomarker scoring in NCTN trials (e.g., sarcomas, NCI-MATCH). Support retrospective AI validation but require standardized identifiers, naming conventions, procedures, and funding. Expand AI use in trial enrollment, biomarker discovery, and harmonization across sites. Advance pathomics by integrating DPI with AI to capture morphological and microenvironmental features and to develop biomarkers. Build large reference datasets, assess bias, and align validation frameworks with regulatory standards.	1,28, 29, 30, 31, 32, 33
LDTs & regulatory context	Validate prognostic/predictive LDTs (e.g., AI for prostate cancer) under CLIA requirements, including analytical performance (accuracy, precision, reproducibility, and reportable range). Distinguish LDT validation from CAP WSI validation for diagnostic workflows. Address adoption barriers (cost, reimbursement). Track evolving FDA/EMA frameworks for AI-based DPI tools. Unlike radiology, where more than 1000 AI/ML devices have received FDA clearance, as of August 2025, only three AI/ML SaMDs in DP have been cleared. This disparity highlights a dataset-driven validation gap rather than an absence of regulatory pathways (e.g., FDA's Medical Device Development Tools (MDDT)).34 The cleared products are Paige Prostate (De Novo, first-of-kind CAD tool), ArteraAI Prostate (De Novo, first-of-kind prognostic tool; initially deployed as an LDT under CLIA), and Galen Second Read (510(k), demonstrating substantial equivalence to Paige Prostate).	2,34, 35, 36, 37, 38, 39, 40, 41
Future directions & national coordination	Sustain US coordination on standardization, federated data exchange, and regulatory-aligned validation. Align with international initiatives (e.g., Bigpicutre, EMPAIA).	8,9,26,27,42, 43, 44

Overview of DPI and AI in precision oncology and the impact of disruptive technology

The digitization of pathology slides has enabled computational image analysis and the use of AI, improved diagnostic accuracy, and thus driving significant advancements in oncology. Advances in computational microscopy, pioneered by researchers like Judith Prewitt, laid the groundwork for modern applications of AI in pathology. More recent breakthroughs in neural networks and deep learning led by innovators, such as Yann LeCun, have propelled the field of computational pathology, enabling automated identification of cellular features, tumor grading, and risk assessment in diagnostics and clinical trials.³⁰ This integration supports streamlined patient screening, clinical trial enrollment, and drug efficacy assessments.

The role of AI in oncology further extends to drug development and evaluating the tumor microenvironment, enabling standardized biomarker assessments that ensure consistent patient treatment across diverse cancer types.^30,45,46 Overall, ongoing tool developments in DPI AI will better position pathologists as central to patient management, in part by enabling detailed tumor analysis and personalized treatment strategies.47, 48, 49 DPI also facilitate large-scale integration of histopathological, genomic, and clinical data, as demonstrated by initiatives such as the 100,000 Genomes Cancer Program. This project leverages WSI and computational pathology to correlate tissue morphology with genomic alterations, supporting the discovery of meaningful—and in some cases actionable—tumor mutations.³³ Such integrative approaches exemplify the potential of precision oncology to transform diagnostics and therapy selection. Similarly, foundational discoveries like EGFR mutations predicting responses to tyrosine kinase inhibitors in non-small cell lung cancer⁵⁰ have catalyzed precision oncology workflows. Recent studies now apply deep learning to digitized hematoxylin and eosin (H&E) slides to non-invasively predict EGFR mutation status, illustrating the role of DPI and AI in advancing molecular diagnostics and stratifying patients for targeted therapies.^33,50

Lessons from digital cytology

Investigators in the field of digital cytology have pioneered the integration of telecytology, WSI, and AI-assisted Pap test screening, which together have significantly improved diagnostic accuracy and operational efficiency. Telecytology enables remote evaluations and real-time consultations, particularly valuable for procedures such as rapid onsite evaluations, which improve specimen triage and thereby improving care in regions lacking pathology specialists. Whereas telecytology offers clear benefits, it requires skilled cytologists and specialized training to manage technical and logistical demands effectively.^51,52 WSI enables remote review and AI-supported analysis, but at the same time, introduces challenges due to cytology's three-dimensional cell structures, requiring techniques such as Z-stacking or volumetric scanning for extended depth-of-field imaging with cellular structures in focus. Recommendations from the American Society of Cytopathology underscore the need for validating digital cytology systems with comprehensive studies reflecting clinical diversity.^53,54 FDA-approved AI-assisted Pap test systems have improved sensitivity and productivity in screening, proving added value of deploying AI in routine cytology practice.^55,56 Key lessons include the importance of skilled personnel, technological reliability, and validation standards, which are critical for scaling DPI in clinical settings.

Insights from the Molecular Analysis for Therapy Choice (NCI-MATCH) precision medicine clinical trial (NCT02465060)

The NCI-MATCH trial, a precision medicine initiative, underscored DPI's role in supporting patient-tailored cancer therapies. This trial matched patients to targeted treatments based on specific genetic mutations, enrolling nearly 6000 participants and demonstrating the feasibility of large-scale precision-driven trials. The requirement for fresh biopsies along with centralized molecular testing and rigorous QCs led to a 93 % success rate in tumor gene sequencing, highlighting the importance of centralized processing and standard protocols in managing diverse clinical sites.^29,57

The implementation of DPI in NCI-MATCH enabled centralized specimen review, QA, and standardized biomarker assessment, although it was resource-intensive. Whereas AI tools were not deployed in real-time during the trial, the availability of digitized slides created opportunities for retrospective AI-driven analysis, highlighting DPI's value in ensuring data uniformity, facilitating specimen QC, and supporting future biomarker research and validation.⁵⁷

The role of DPI in clinical trials is increasingly significant for establishing standardized testing protocols and centralized biomarker testing infrastructure. As highlighted in workshop presentations, DPI is being operationalized in active National Clinical Trials Network (NCTN) trials, especially in rare cancers such as sarcomas, where centralized expert review of small, heterogeneous cohorts ensures diagnostic accuracy and harmonized biomarker scoring across sites. WSIs further support prospective QA/QC (e.g., focus, artifacts, and resolution), consistent acquisition parameters, and uniform application of scoring rubrics (e.g., PD-L1, TILs), thereby reducing variability and improving reproducibility of histological endpoints. In practice, DPI creates a durable digital archive that not only supports primary trial endpoints but also enables retrospective AI-driven biomarker validation and QC, as demonstrated in NCI-MATCH. Unlike radiology, pathology faces dual interoperability challenges: (1) syntactic interoperability—exchanging tiled, multi-resolution WSIs across heterogeneous vendor formats and APIs and (2) semantic interoperability—harmonizing annotations, labels, and metadata. Workshop discussions highlighted how DICOM adoption can alleviate these challenges by reducing the burden of multi-vendor scanner/IMS integration, whereas standardized de-identification protocols ensure secure cross-site data sharing. Using DICOM together with standardized vocabularies and metadata models addresses both challenges, enabling scalable, multi-institutional trial workflows.^1,4,29,57, 58, 59

Key takeaways for DPI applications include the necessity of standardized testing protocols, centralized biomarker testing infrastructure, investment in advanced technology, and collaboration across research sites to maintain quality and efficiency.^60,61

Standardization and data sharing in radiology as a model for DPI: Lessons from radiology and The Cancer Imaging Archive (TCIA)

Radiology's successful integration of standardized protocols offers a robust model for DPI. Central to this success is the adoption of the DICOM WSI, which facilitates interoperability by allowing images from various devices and manufacturers to be stored, retrieved, and analyzed across platforms.⁶² This approach overcomes diversity across imaging modalities and vendors by enabling relatively seamless integration through standardized data, reducing proprietary limitations and dependencies. DICOM's dual focus on syntactic and semantic interoperability—standardizing data formats and transmission protocols (syntactic), and ensuring consistent interpretation of codes and descriptions across systems (semantic)—enhances cross-institutional data comparison, image fusion, and AI applications, all essential for expanding research and clinical workflows.^13,63 Radiology also prioritizes exam procedure normalization and systematic coding practices, using resources such as SNOMED CT⁶⁴ and LOINC⁶⁵ for consistent naming and benchmarking. Image annotation frameworks, such as DICOM Structured Reporting (DICOM SR) and Segmentation (DICOM SEG), further optimize data discoverability and reuse.^17,18 These practices underscore the need for DPI to adopt similar open standards and standardized coding to support AI-ready data management.^66,67

The Cancer Imaging Archive (TCIA)²⁷ provides a model for data sharing that extends radiology's standardization. TCIA, a repository for de-identified imaging data, hosts over 200 curated datasets, drawing approximately 20,000 monthly visitors and enabling over 200 TB of data downloads. TCIA supports extensive AI-driven research, algorithm development, and cross-institutional collaboration.⁶⁸ By including diverse imaging data from X-rays to pathology slides, TCIA promotes interoperable, open-access data, contributing to AI-driven research and algorithm development with significant publications and patent applications.^22,23 One example of TCIA's impact is the use of histopathology WSI from the The Cancer Genome Atlas (TCGA) glioblastoma collection, where researchers successfully developed AI models capable of distinguishing glioblastoma from healthy brain tissue with high accuracy.⁶⁹

Investigator needs in DPI and AI-driven research

To support DPI effectively, it is essential to understand and address the needs of investigators in the field of AI-driven research. Investigators working with DPI and AI-driven research require high-quality, structured data to drive precision and advance personalized therapy. Effective AI-driven pathology research requires high-quality, structured data. Essential elements include:

• •High-resolution, well-validated imaging datasets.
• •Standardized metadata for multi-modal analysis.
• •Cloud-based platforms for scalable AI validation.

Emerging technologies such as multiplex immunofluorescence imaging and the concept of Cancer Digital Twins—virtual computational models that attempt to replicate an individual patient's tumor biology or certain aspects thereof—are transforming pathology images from static visuals into rich computational data, potentially enabling predictive modeling of therapeutic response, hypothesis-driven research, and personalized therapy in clinical trials.^70,71

Multi-modal imaging capabilities, such as those offered by cyclic immunofluorescence, provide molecular-level details on tumor morphology. Yet, challenges in validating antibodies and ensuring image processing consistency necessitate specialized solutions, for example, Minerva, which facilitates multiplex imaging review.⁷² Access to platforms that integrate multiplex data with broader clinical insights, such as cBioPortal, enables researchers to explore tissue-specific biomarkers effectively. Notable applications include targeting BRAF V600E mutations in papillary craniopharyngioma⁷³ and studies on ovarian cancer treatment responses, cellular cycle metrics, and immune cell differentiation in lung adenocarcinoma.^74,75 Moving multiplex imaging into clinical practice requires simplified assays and reliable WSI capabilities that are compatible with existing clinical workflows. Early commercial methods, such as RareCyte's Orion, combine molecular and H&E images to reveal spatial tumor biomarkers, though the low research volume remains a hurdle for wider clinical adoption.

The Multi-Omics Multi-Cohort Assessment (MOMA) platform developed using data from NCI-sponsored studies including TCGA, and the Prostate, Lung, Colorectal and Ovarian cancer screening trial further underscores the need for integrated high-quality data. MOMA leverages multi-cohort datasets to predict molecular profiles and prognosis in colorectal cancer patients, proving generalizable across diverse demographics and image digitization methods.⁷⁶ Although MOMA models demonstrate high accuracy in predicting multi-omics profiles and survival rates, AI applications in clinical settings still face challenges related to contextual nuances such as artifacts, lab contexts, clinical correlations, and scanner variability.⁷⁷ Addressing these challenges may require the integration of implementing out-of-distribution detection, multi-modal learning³¹ and considerations for human-in-the-loop systems.³²

WSI, tumor biomarkers and foundation models. WSI is central to DPI, capturing rich information on tumor morphology, immune cell infiltration, and tissue architecture that informs cancer diagnosis, prognosis, and treatment response prediction. A prominent example is the analysis of tumor-infiltrating lymphocytes (TILs), which has shown prognostic value across multiple cancer types.^46,47 Recent advances in AI, particularly weakly supervised and self-supervised learning, allow model development using minimal labeled data, making it feasible to scale analysis across large WSI datasets. These methods underpin the creation of foundation models, which can later be fine-tuned for downstream tasks such as segmentation, classification, or molecular prediction. Whereas these approaches are promising, especially in reducing the burden of annotation, their clinical effectiveness still requires rigorous validation. This underscores the ongoing need for high-quality, annotated WSI datasets, and interoperable tools for spatial mapping and multiomics integration.

Real-time pathology assessment tools are invaluable in clinical settings, especially in critical, time-sensitive scenarios like surgery. The AI-based Cryosection Histopathology Assessment and Review Machine system enables rapid identification of malignant cells in cryosection samples during brain cancer surgery. Its high predictive accuracy for mutation status and histological grading illustrates AI's potential in high-stakes decision-support and emphasizes the necessity for validated AI models that ensure reliability across varied clinical conditions.⁷⁸

Validating AI models in clinical trials. The reproducibility and reliability of AI in medicine hinge on rigorous validation. Research underscores the importance of explainable AI, multi-institutional validation, and clinical trial data in tackling AI's reproducibility challenges. Validation frameworks incorporating diagnostic, prognostic, and predictive tools ensure that AI technologies provide accurate clinical decision-making. Examples like deep phenotyping—where AI maps tissue features for prognostic markers—and the ECOG 5163 trial, which uses radiomic data to forecast treatment responses, demonstrate the value of validation across diverse clinical environments.^79,80

DPI in biobanking for clinical trials

DPI integration into biobanking for clinical trials reshapes traditional biobanking workflows, introducing complexities in IP management, data acquisition, storage, and distribution. Clearly, DP extends beyond imaging, encompassing data interpretation and management, which demands robust infrastructure and strategic handling of IP rights. This includes navigating initial setup costs, maintaining data interoperability across scanner systems, and ensuring compliance with data security and privacy standards.²⁸ In the context of NCI clinical trials, banked DPI data have significantly simplified the central pathology review process. It enables remote, multi-pathologist reviews, eliminating the need for physical slide transfers. Using open-source tools, pathologists can directly collect trial-specific data, though consistent practices are needed for image annotation and discrepancy management. Typically, image data are housed at the imaging site or at statistical centers for redundancy, whereas clinical data are accessible either directly or via the NCI CTEP Data Archive.^59,81

High-quality DPI acquisition and management. Ensuring consistent, high-quality DPI across clinical trial sites poses significant challenges. One primary challenge is the reluctance of many clinical sites to release diagnostic slides for centralized scanning. Even with assurances of prompt return, many institutions are hesitant. Additionally, a significant number of clinical sites lack the infrastructure to scan slides, particularly smaller community hospitals involved in multi-center trials. The diversity in scanning platforms and the absence of standardization further complicate the process. Unlike radiology, DP has not benefited from universally accepted standards such as DICOM, even though there has been a DICOM WSI standard for several years, resulting in variability in scan quality, magnification, and de-identification practices. This variability poses a problem for the consistency and usability of the scanned images.^59,82 A centralized DPI portal integrated with the biospecimen submission system such as the Alliance's Digital Pathology Imaging Portal (DPIP), helps standardize image acquisition, enforce naming conventions, and ensure logical links between images and biospecimens.⁵⁹

Image de-identification and QA. Ensuring that images from biospecimens are free of Protected Health Information (PHI), and personally identifiable information (PII), is crucial to maintaining patient confidentiality. Many clinical sites inadvertently submit images containing PHI, necessitating robust QA and de-identification workflows. DPIP's structure supports this by enforcing protocols that link images to non-identifying patients and study information. Image de-identification is further discussed in a later section.

QA in pathology refers to the system-wide protocols and standards that are put in place to ensure consistency, reliability, and accuracy across all stages of tissue processing, diagnostic interpretation, and data generation. This includes activities like standardized operating procedures, personnel training, accreditation compliance (e.g., College of American Pathologists (CAP), Clinical Laboratory Improvement Amendments (CLIA)), and audit trails. Current CAP WSI validation guidance and CLIA requirements emphasize verification/validation, reproducibility, QA/QC, and traceability of DPI for diagnostic use. Whereas these guidelines do not explicitly mandate or reference DICOM for DP, adopting the DICOM WSI standard is highly complementary: It provides uniform image exchange and metadata capture, facilitates multi-site validation and proficiency testing, strengthens audit trails, and supports scalable interoperability. Furthermore, the FDA references DICOM as a “Recognized Consensus Standard” for DP systems that vendors may reference in their clearance submissions.¹⁶ In DP, QA encompasses scanner calibration routines, adherence to image acquisition standards, and validation of software tools and AI algorithms before deployment. QA measures are essential to verify image quality, including stain clarity, focus, and required formats. These structured workflows allow DPI systems to provide comprehensive, high-quality datasets for secondary research. Other QA measures may relate to ensuring algorithm interoperability with data from multiple sites in a clinical trial, which could be done through a federated learning approach.²⁵

Preparing DPI for AI applications. To support AI research, DPI data must be consistently high-quality, well-annotated, and standardized. Variability in imaging devices and formats across sites complicates AI applications. Implementing structured protocols for image capture, formatting, and annotation is essential to create AI-ready datasets. Multi-level annotation, including discrete data and visual markup, enhances DPI datasets, enabling AI algorithms to perform tasks such as segmentation and prediction with greater accuracy. Including imaging objectives in trial protocols also supports dataset preparation for AI applications, improving prognostic model validation and accuracy.⁸¹

Image format and annotation standards. Whereas standards such as DICOM for pathology imaging do exist, widespread adherence to these standards remains limited. This lack of consistent adoption across institutions and vendors hinders interoperability and complicates cross-site data integration. Although DICOM offers significant benefits for standardization, adoption has been slow, largely due to limited vendor support rather than inherent complexity. Efforts are underway to establish standard file-naming conventions and metadata extraction protocols across imaging platforms, enhancing compatibility and reducing barriers to analysis. Improved indexing of image databases would streamline image access and allow investigators to leverage a larger, more standardized dataset for research.⁸¹

Strategies for long-term storage, sharing, and retrieval. Cloud-based object-oriented storage is increasingly being considered for WSI archival and exchange. Object-oriented storage, in which images are treated as discrete addressable objects, scales efficiently to multi-petabyte datasets and aligns with DICOMweb tile-access protocols. This architecture supports distributed access across trial sites and facilitates AI/ML workflows, whereas compliance with CAP and CLIA requirements for validation, auditability, and data integrity remains essential. Hybrid deployments—integrating local storage for primary diagnosis with cloud storage for research and secondary use—are emerging as a practical approach for clinical trials. Long-term storage of DPI data presents both logistical and financial challenges due because of the large file sizes involved.⁵⁹ Cloud-based storage options, such as Amazon Web Services, Google Cloud Platform, and Microsoft Azure, simplify data upload and retrieval, whereas potential collaborations with repositories like TCIA²⁷ and the NCI Imaging Data Commons (IDC)²⁶ can provide sustainable long-term storage and distribution. Establishing clear policies for data storage duration and developing scalable retrieval systems ensure data remain accessible for future studies.⁸³

Financial support and prioritization. Given limited funding, prioritization is essential to support comprehensive DPI efforts across clinical trials. Criteria for prioritization include trials with high scientific impact, novel applications, or potential for AI-driven research.⁵⁹ For example, trials investigating immune checkpoint inhibitors in rare tumor types or genomically directed trials designs such as the NCI-MATCH have been prioritized due to their potential to generate high-value imaging and molecular data.⁵⁷ This strategic focus ensures that resources are allocated efficiently, supporting imaging for trials with the greatest potential to drive research and discovery.

IP and image sharing. Clear data-sharing agreements and defined IP rights are essential for maximizing the research value of DPI data. Images, treated as data, are governed by Data Use Agreements that outline permissible uses and restrict unauthorized sharing. Including DPI objectives proactively in clinical trial protocols, as in NCI's NCTN breast cancer trial S-2212 (SCARLET),⁸⁴ addresses ownership and access issues, facilitating open access while ensuring data integrity. Standardized agreements and clear guidelines for integrating clinical and genomic data enhance DPI data's value, supporting both collaborative and independent research.⁸³

Standardization, de-identification, and validation in DPI

Standardization in DPI enhances interoperability and re-usability of data. In the same manner as is well-established for radiology, DICOM for pathology facilitates image archival, exchange, and exploration of image metadata, addressing clinical research needs and improving efficiency. The NCI IDC is an example of a public archive in full DICOM compliance, providing radiology, pathology, and annotations in the DICOM standard.²⁶

Although some consider DP to lack standardized practices, David Clunie, editor of the DICOM standard, highlighted the existence of a DICOM WSI standard developed around 2010 with development pursued under the guidance of DICOM Working Group 26.^13,14 DICOM, already promotes interoperability by enabling consistent data exchange across systems, images, annotations, and protocols, and defines comprehensive metadata for identifying and describing patients and specimens and their preparation as well as acquisition techniques for brightfield, fluorescence, and multiplex microscopy techniques, which supports both clinical and research applications, with open access and extensive support from open-source tools. Pathology-specific extensions for very large numbers of annotations from computational pathology are also defined in DICOM.⁸⁵ Despite relatively slow adoption in pathology, DICOM WSI could benefit DP significantly by leveraging and re-using existing scalable enterprise-wide infrastructure for storage, security, and business continuity, helping bridge gaps in clinical practice rather than requiring localized pathology-specific proprietary implementations, allowing users to select appropriate mixed vendor best-of-breed solutions. Despite DICOM's benefits, its limited vendor support poses continued challenges to achieving full interoperability in DPI. Any standard may be better than no standard, however.^24,86,87

Open-source tools for WSI de-identification. To share WSI data securely, de-identification is necessary to comply with regulations like the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation. De-identification in DP is complex and extends beyond redacting structured metadata. PHI can also appear in filenames, embedded image metadata, macroimages, and scanned slide labels. These potential leak points require careful handling.^88,89 Data sharing through public (TCIA and IDC) or private (Digital Slide Archive, and DSA) archives necessitates de-identification of images before sharing them. Whereas DSA itself does not perform the de-identification, it supports integration with external tools such as WSI DeID—a de-identification algorithm that removes PHI/PII from WSIs—and visualization tools like HistomicsUI/TK.^24,90 The platform's ImageDePHI pipeline⁹¹ incorporates human oversight for QA, ensuring no information is compromised and avoiding potential artifacts that can arise from machine processing alone.^11,13,24

Validation of DP and AI systems. Validating DP and AI systems is essential for clinical reliability and regulatory compliance. CAP issued guidelines in 2021 for validating WSI systems in diagnostic pathology, including using sample sizes of at least 60 cases to establish diagnostic concordance between digital and glass slides, with a minimum target concordance of 95 %.¹¹ This requirement applies specifically to WSI platforms for primary sign-out and should be distinguished from AI model validation. AI validation generally requires much larger datasets—often in the thousands of slides—to ensure robustness, generalizability, and mitigation of demographic or disease-specific biases. Beyond CAP, the FDA provides technical guidance²⁰ and has cleared several DP platforms that incorporate AI-assisted tools, including scanners and viewers for clinical use; however, the broader regulatory framework for adaptive or continuously learning AI diagnostics is still evolving.⁹² Similarly, the European Medicines Evaluation Agency has issued regulatory guidance for in vitro diagnostic medical devices, including WSI systems and image analysis software.³⁸ In clinical labs, laboratory developed tests (LDTs) are frequently applied to DPI and must be developed and validated using CLIA standards to demonstrate analytical validity before clinical use. Validation involves establishing concordance between glass and digital workflows, typically following CAP guidance for WSI validation (≥60 representative cases with ≥95 % agreement, plus additional cases for immunohistochemistry or special stains, and use of washout periods to reduce recall bias). Under CLIA, LDT validation must also document accuracy, precision, reproducibility, reportable range, and analytical sensitivity/specificity. When LDTs incorporate AI/ML tools for biomarker quantification or prognostic modeling, validation extends beyond CAP's minimum case requirements and generally requires larger, multi-site datasets, discordance analyses, and bias assessments to ensure robustness and generalizability. All validation studies must be documented and approved by the lab director, with ongoing monitoring and revalidation as DPI platforms, staining protocols, or algorithms evolve.^2,93 ArteraAI initially deployed its multi-modal AI-based prognostic assay for localized prostate cancer as a LDT under CLIA certification. Following multi-institutional validation studies, the assay received FDA De Novo authorization in August 2025 (DEN240068), representing the first Software as a Medical Device (SaMD) clearance for an AI-driven prognostic test in DP.^35,36 In addition to ArteraAI, only two other AI/ML SaMDs in DP have FDA clearance to date: Paige Prostate (De Novo, first-of-kind CAD tool)⁴⁰ and Galen Second Read (510(k)), demonstrating substantial equivalence to Paige Prostate.⁴¹ These examples highlight both the promise and the challenges of moving AI tools through rigorous validation to regulatory clearance.³⁴ Best practices for machine-learning models include real-world data, ensuring rigorous validation, and addressing demographic and disease-specific biases.^21,35, 36, 37

Hardware, software, and data management solutions in DPI

Digital (WSI) scanners are essential in DP for supporting diagnosis, teleconsultation, and image analysis. Scanners capture and manage digital slides through a combination of hardware components—optics, lenses, cameras, and robotic systems—alongside specialized visualization software.⁹⁴ Scanners create WSI by capturing thousands of image tiles and stitching them together, enabling pathologists to view slides digitally. Other less expensive or technically complex systems can use a live digital feed on top of a microscope and driven by the viewer to generate large area digital files (at multiple resolutions) live, such as Microvisioneer ® system.⁹⁵

Image quality is influenced by both traditional microscope elements and digital capture processes, including post-processing and display.^96,97 However, WSI offers advantages over traditional microscopy, improving portability, sharing capabilities, and digital analysis potential. Studies demonstrate that WSI is non-inferior and even outperforms traditional microscopes in certain diagnostic tasks.⁹⁸

Scanner performance varies depending on factors such as quality, illumination, scanning speed, scanning capacity, software capabilities, and advanced features including autoloading and continuous slide loading.⁹⁹ Typical WSI scanning involves capturing tiled images at multiple resolutions using real-time focusing, with magnification levels commonly at 20× and 40×, depending on diagnostic requirements (with optional z-stacking for depth of field) to create a pyramidal multi-resolution structure for digital navigation.^100,101 Adoption of scanners in clinical settings remains slow due to cost and infrastructure demands, including availability of trained staff and necessary workflow adjustments. Initial expenses include acquiring scanners, upgrading network capacity, and data storage, cost of software integration, with additional ongoing costs for maintenance.¹⁰² Beyond traditional slide-based imaging, Slide-Free Microscopy techniques now enable ex-vivo analysis of un-sectioned tissue, bypassing steps like fixation and staining.¹⁰³ Techniques such as confocal and light-sheet microscopy hold promise, though they are costly and complex. Fluorescence Imitating Brightfield Imaging technology, further enhances imaging from unprocessed tissue, using autofluorescence back illumination to mimic H&E images while preserving tissue integrity.¹⁰³

Quality control in WSI. QC is critical for reliable DP workflows. QC refers to specific checks and tests performed during or after a process to identify defects or deviations. For example, verifying section thickness, confirming stain quality, assessing digital image resolution and color fidelity after scanning, and ensuring that images meet the pre-defined parameters for analysis. In AI-based pathology, QC may also involve verifying that input images conform to model requirements and performing visual review of outputs to catch false positives or artifacts.

Effective QC for WSI involves pre- and/or post-scan checks, verifying tissue detection, focus, and image integrity. QC processes benefit from trained histology technicians and consistent practices, including review of macro-level images to flag issues.¹² AI-enhanced QC systems combined with human oversight, improve efficiency by identifying slides needing rescanning, reducing manual QC burdens.¹⁰⁴

DICOM and software development for DP. The DICOM standard, especially the Whole Slide Microscopy Image (DICOM WSI),¹⁰⁵ introduced by DICOM Supplement 145 and subsequently updated since, supports DP by embedding metadata that reduces transfer errors and ensures data integrity.^106,107 Widely used across medical fields, DICOM WSI is instrumental in image acquisition, storage, retrieval, and analysis. The DICOMweb network protocol allows user-friendly, portable web-based access to a virtual microscopy experience as well as access to tiles for computational pathology training and inference. To achieve full interoperability adopting DICOM WSI requires users to thoroughly understand the complexity of the workflow as well as to require support directly by the scanner vendors rather than depending on conversion from existing proprietary formats.⁸⁶ Integration with the AP-LIS (laboratory information system) to obtain appropriate metadata during scanning is essential and standards exist to support this such as the Integrating the Healthcare Enterprise Digital Pathology Image Acquisition Profile¹⁵ which uses messaging standards from Health Level Seven (HL7), a widely adopted framework for electronic health information exchange. Vendor DICOM WSI implementations of images and annotations are tested and evaluated at Connectathon events.¹⁰⁸ Relevant information for digitally scanned slide includes identifying and descriptive metadata about the patient, case, specimen (including staining), and the mechanics of the slide acquisition (such as magnification). DICOM, HL7 version 2 (V2), and HL7 FHIR provide mechanisms for slide scanners, IMS, and LIS to exchange this information. It is possible to encode DICOM WSI from the scanner with very limited metadata (so-called “naked” DICOM files). It is preferable for the scanner to use the scanned barcode identifier to query the LIS to populate the DICOM metadata with patient and specimen information. The IMS may directly query the LIS or the EHR using HL7 V2 or FHIR to retrieve additional information relevant to for interpretation of the slide.¹⁰⁹

AI integration in pathology software aids workflows through tasks like ground-truth generation, training set selection, and model deployment. Annotation tools¹⁰⁸ within WSI viewers allow pathologists to efficiently label slides, with software supporting metadata attachment, multi-annotation tracking, and remote access. The AI workflow encompasses training, testing, and deployment, with model outputs displayed as heatmaps or other aids for diagnostic evaluations.^19,110

Data management. The NCI IDC²⁶ serves as a cloud-based repository for open access to cancer imaging data, including WSI, radiology, and fluorescence images. IDC is the repository used by the NCI Childhood Cancer Data Initiative, among other programs, for distributing DPI.¹¹¹ IDC standardizes all images in DICOM format, most of which are available under a commercial-friendly license, simplifying access for research and collaboration. Beyond archiving, IDC harmonizes all the images and analysis results into a standard DICOM representation and is actively engaged in enriching the hosted data using AI-based curation approaches. IDC's use of cloud-based services for data streamlines integration with the emerging AI capabilities and aims to simplify cloud-based analysis of the data.⁴² TCIA complements IDC by focusing on the de-identification and distribution of curated medical images, supporting DPI and AI model training. TCIA's expertise in curation ensures data quality and accessibility for various applications.¹¹²

In addition to US-based initiatives such as the NCI IDC and TCIA, international programs are advancing DPI infrastructure at scale. The Bigpicutre project, supported by the European Union, is establishing a federated, pan-European repository of DPI to accelerate AI training, validation, and regulatory engagement.⁴³ Likewise, Germany's EMPAIA initiative is building a reference architecture and certification framework to enable safe, interoperable AI deployment in DP practice.⁴⁴ Together, these efforts complement US initiatives and highlight the value of global public–private partnerships and standards harmonization in advancing DPI-AI validation and adoption.

Enterprise platforms for image management and analysis

To manage the complexity of diverse imaging modalities, an enterprise data science platform is crucial to support identification, data management, and workflow deployment, enabling high-throughput research. These platforms utilize open-source tools like OMERO web, RStudio, and JupyterHub, and rely on cloud or high-performance computing for simultaneous multi-experiment processing, essential for large-scale DP research.¹¹³

Developing image data repositories for research. With the rising demand for comprehensive H&E image repositories, particularly for rare cancers, the NCI's Cooperative Human Tissue Network has launched a project to create a major H&E image repository. This repository will provide cloud-based access with AI-enabled tools and standardized annotations, expediting annotation processes and supporting advancements in AI-driven DP.¹¹⁴

Integrating DPI into clinical trials

Whereas DPI was not part of the original NCI-MATCH trial design, its potential value in interpreting complex therapeutic responses was emphasized by Dr Stanley Hamilton and Dr Keith Flaherty during this workshop. Coordinated by ECOG-ACRIN and the NCI, the trial revealed the complexity of response, underscoring DPI's value in refining therapeutic strategies. In the context of NCI-MATCH, DPI was used retrospectively to review tumor slides, sometimes leading to re-classification of tumors in a way that could have influenced treatment arm assignment. This underscores the potential for DPI to complement molecular profiling by capturing biologically relevant features such as TILs, tumor heterogeneity, and cancer cell states.^1,19,20,92

The trial highlighted the limitations of genetic sequencing in predicting responses, showing that genetic complexity, co-occurring mutations, and tumor histology also impact drug efficacy. Both targeted therapies and immunotherapies showed variable effectiveness influenced by cancer cell states and tumor-infiltrating immune cells. Additional studies have identified other biological mechanisms contributing to targeted therapies resistance, such as epithelial-to-mesenchymal transition in EGFR-mutant lung cancer and neuroendocrine differentiation in prostate cancer. These examples emphasize how tumor cell states and microenvironmental features—factors not always captured by genomics or transcriptomics—can impact drug response. To better account for these complexities, pathomics has been proposed as a complementary strategy. By extracting morphological and contextual features from pathology images, pathomics can uncover subtle predictors of response and resistance, helping to refine precision oncology and improve therapeutic targeting.¹¹⁵

In osteosarcoma, automated histological evaluations enabled precise assessments of tumor necrosis, improving diagnosis and outcome predictions.^116,117 Similarly, digital imaging in rhabdomyosarcoma has allowed outcome predictions based solely on histological analysis, demonstrating how advanced imaging can enhance precision in pediatric cancer treatment¹¹⁸ while controlling diagnostic costs.⁵⁸

Developing AI biomarkers through DPI is a promising direction for clinical trials, enabling applications in patient stratification, disease progression, prediction, and response assessment. The development process typically begins with concept discovery and feasibility studies to ensure that the clinical question and dataset are appropriate.¹¹⁹ Following this, model building focuses on minimizing bias, setting prediction thresholds, and ensuring out-of-domain (OOD) recognition to safeguard reliability. Multi-site and prospective clinical trials are then used to validate model efficacy in realistic clinical settings. The final steps involve regulatory compliance and integration into standard clinical workflows, facilitated by early collaboration with regulatory bodies to clarify requirements and streamline approval. Addressing potential biases and enhancing model safety strengthens the accuracy of AI-powered biomarkers. Importantly, early engagement with reimbursement frameworks can encourage a smooth transition from research to clinical application.¹²⁰

Intellectual property (IP) issues in DPI

IP rights in general. The US property system encompasses real, personal, and IP. IP, though intangible, is a valuable creation of human intellectual activities, primarily protected by patents, trademarks, trade secrets, copyrights, and/or contract laws. IP ownership rights include possession, exclusion, use, transfer, and enforcement. IP owners typically use licenses in various formats to permit specific access and use by others under set conditions.¹²¹

IP types in DPIs. DPIs usually involve several IP types: clinical data (or databases if it is a structured collection), information, and copyrighted images. Clinical data and information are typically protected by contract law, whereas image copyrights are automatically assigned upon creation, often residing with the authors or institutions that generate the images.¹²¹

IP considerations in DPIs. IP rights in DPIs stem from initial acquisition through creation, subsequent licensing, or transfer of ownership. Key aspects include determining DPI’s ownership and checking existing contractual or legal restrictions on licensed uses. Notable cases such as Moore v. Regents of the University of California¹²² and the Henrietta Lacks settlements underscore the limited control patients have over their biospecimens, including data and information generated therefrom, as well as the importance of informed consent in downstream uses of such materials and information.^121,123

Patient rights in biospecimens and DPIs. Patients do not retain ownership of their excised tissues and cells once removed, as established in Moore v. Regents of the University of California (1990), which ruled that individuals do not have property rights over their discarded biological materials used for research. Control over the use of DPIs derived from those biospecimens is primarily governed by informed consent and privacy laws such as the HIPAA.^122,124 Any use or sharing of DPIs must align with the scope of patient consent and institutional review board approval. Furthermore, compliance with HIPAA requires de-identification of PHI before secondary use or broader data sharing, as outlined in the HIPAA Privacy Rule.^121,125

Biobanking and data management. Biobanks nowadays play a central role in managing DPIs, using open-source or controlled access modes to balance accessibility and IP law compliance. Examples include the NCI IDC and TCIA, which provide public access under specific licenses and agreements set up by DPI's owners or providers.¹²¹

Takeaways and recommendations

Investment in DPI infrastructure, along with standardized protocols and IT staff, is critical to ensure reliable data management, image acquisition, and secure storage across clinical, diagnostic, and research applications.

Successful deployment of DPI requires optimization of workflow, both before digitization and after slides are scanned, to enable secure remote access. Integration of the IMS with a LIS should be prioritized to streamline operations.

Adoption of the DICOM WSI in NCI clinical trials, similar to radiology, is recommended to enhance interoperability, allowing seamless data exchange across imaging platforms, research sites, and from a data science perspective, across all DICOM objects (radiology, pathology, radiation therapy, and reports).

Open-source tools platforms, such as the DSA platform designed for data management, visualization, and analysis, enhance data-sharing capabilities by securely de-identifying DPI. Alternatively, federated learning approaches, which allow AI models to be trained across multiple institutions without exposing patient privacy, represent another approach that could be useful in certain circumstances.

Data-sharing platforms like the IDC and TCIA, which comply with the HIPAA privacy rule, facilitate centralized storage of, and access to, de-identified medical and pathology images, crucial for collaborative AI research.

CAP guidelines provide a foundational approach for validating WSI systems for primary diagnostic use, ensuring diagnostic safety and clinical reliability, thereby promoting the adoption of WSI in clinical practice.

Comprehensive validation frameworks that simulate real-world clinical settings are critical for the successful deployment of DPI and AI technologies in pathology, supporting diagnostic precision and advancing research in oncology.

Integrating DPI into biobanking and clinical trials requires effective collaboration between trial investigators, biobank managers, and statisticians. Developing centralized systems, standardizing procedures, standardized naming conventions and addressing funding needs are crucial for biobanks to maximize the accessibility and research potential of DPI.

For dealing with IP generated via DPI, inventors should work with deidentified digital data DPIs, employ agreements to control downstream data sharing and uses, and enforce rights diligently. Users must keep detailed records of such agreements, understand permissions and restrictions, and separate data arising from different sources to avoid downstream IP conflicts.

In clinical trials such as NCI-MATCH, DPI supported centralized pathology review, ensuring consistent evaluation of specimens across sites. Although not used in real-time to refine patient selection or biomarker standardization, the availability of digitized slides enables retrospective analyses that may inform future biomarker development and stratification strategies.

Prognostic and predictive LDTs using DPI are available (e.g., deep learning algorithms), but concerns about return on investment limit their clinical adoption. Institutional support and sustainable reimbursement opportunities could help, whereas evolving regulations may facilitate the shift from research to clinical practice.

Pathomics (i.e., the combination of DPI with AI to extract quantitative features from WSI of tissue sections) is a promising method to capture morphological and microenvironmental features within tissues beyond what genetic sequencing alone provides, aiding predictions of therapeutic responses and uncovering complex tumor interactions within the tumor microenvironment, while identifying AI-driven digital biomarkers.

AI-driven digital biomarkers for diagnosis, prognosis, and treatment response prediction hold promises for optimizing patient-specific treatment strategies. Ensuring early identification of confounders and biases in AI model development is critical for improving clinical reliability.

To enhance AI model reliability and ensure patient safety, integrating uncertainty estimation and OOD detection is essential, allowing models to recognize and appropriately manage unfamiliar data.

Multi-disciplinary collaboration plays a vital role in establishing essential reference datasets that strengthen AI validation and inform regulatory decisions. In addition, publicly available datasets contribute to improved accuracy, reproducibility, generalizability, and interoperability in DP.

The evolving regulatory and reimbursement landscape can facilitate the transition of prognostic and predictive LDTs using DPI from research to clinical practice but also presents challenges. Stringent regulations may hinder AI adoption by adding complexity, cost, and lengthy validation processes. A balanced approach is needed to ensure patient safety while fostering innovation, with regulatory processes that uphold scientific rigor while streamlining approval for AI-based DP tools.

Several workshop themes—particularly around standardization, data sharing, and regulatory-aligned AI validation—align directly with national efforts (e.g., PIcc Alliance for Digital Pathology and FOCR Digital PATH Project) through its Alliance for Digital Pathology which seeks to standardize DP processes, promote synergistic regulatory science efforts as it applies to DP and accelerate innovation for patient care as well as broader ongoing international initiatives aimed at cross-domain data sharing across radiology, pathology, and surgical imaging. Sustained investment in such programs will be critical to achieving scalable, interoperable infrastructures, and clinically integrated solutions for DPI and AI in cancer research and clinical trials.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article.