3.1.

Assessing the Clinical Utility of AI in Pathology

As stated before, digitization serves as a necessary first step toward enabling AI within AP. Within an appropriate information technology (IT) infrastructure, whole slide imaging, paired with computational pathology and AI tools, has the potential to greatly automate diagnostic pathology workflows and create enhanced pathology reports (Fig. 1). Marked differences between traditional manual AP lab processes and AI-enabled workflows include the following:

Fig. 1

Clinical impact of digital pathology, computational pathology, and AI on the traditional AP workflow: (a) traditional AP workflow and (b) AI-enhanced AP workflow. Dark blue denotes manual tasks, and light blue denotes AI-enhanced/automated tasks. Note that the AI-enhanced workflow assumes the application of many different AI tools used in tandem, with both the individual tools and the multi-staged clinical workflows all validated for clinical use by the pathology practice.

Of note, each of the above examples’ clinical utility may not be applicable to all pathology practices. Instead, each specific intended use of AI tools should be assessed to determine not only their feasibility within a specific laboratory practice but also how one would best validate the tool for clinical use for specific patient populations and pathology workflow methods.

Although the workflows presented in Fig. 1 represent an optimized, future-state view of AP, current-state AI algorithms under development or ready for deployment for AP labs are more focused in nature. In general, these algorithms can be functionally classified into four major categories:

Although most individual algorithms fall primarily into one of these classifications (e.g., Fig. 2), many AI models are actually combinations of multiple algorithms that tackle different tasks. For example, an AI model that predicts the Gleason grade for prostate cancer may have algorithms that first detect the anomalous areas on the WSI as putative tumor and then characterize the glandular structures into the known patterns of prostatic adenocarcinoma. Finally, the model will then quantify the patterns to estimate the Gleason score. Alternatively, a similar model can be developed, perhaps using the same building blocks, to predict patient outcome directly without a Gleason grading intermediary.²⁵

Fig. 2

Examples of three classes of AI employment in AP: (a) quantitative immunohistochemistry, in which cell nuclei are detected, and biomarker staining intensity is quantified; (b) heat map overlaid on a low power digital image of H&E stained tissue to direct the pathologist’s gaze to the presence of a histologic feature of interest; and (c) slide-level characterization of a prostate biopsy based on integration of regional classifications.

One should also consider that algorithms can be classified based on their clinical utility itself, i.e., the justifications given to how they will add value to patient care. Although there can be multiple reasons to justify incorporating an algorithm into a pathology practice, those most commonly used today in AP concern optimizing efficiency and improving quality. Criteria for optimizing efficiency include implementing algorithms to decrease turnaround time, triage patient cases by acuity, reduce the need for manual intervention, and eliminate tedious activities. For improving quality, the focus is typically on creating more reproducible histopathology outputs used to standardize pathology synoptic reporting or to inform predictive models. The net effect of these modifications may be a reduction in costs, which itself can be a major source of clinical utility for labs looking to make the leap to going digital.

Table 1 summarizes the advantages of AI in pathology and limitations that should be addressed to optimize its clinical utility. Notably, the quality of AI alone does not automatically translate into superior diagnostic quality or greater efficiency. One should note that there are use cases in which even a perfectly-performing algorithm provides little to no benefit for many laboratories or carries no cost savings, efficiency gain, or quality improvement. Ultimately, one must identify those use cases and deployments in which AI can markedly improve current state workflows and provide optimal clinical utility.

Table 1

Potential advantages and limitations of AI use in anatomic pathology.

3.2.

Employing AI in Pathology Workflows

In practice, one of the first considerations for employing AI in the laboratory is understanding how to integrate it into clinical or operational workflows. These typically take the form of operational algorithms that assist with data collection, case triage, screening, and QC and assistive algorithms that contribute through human-AI interaction. For example, an operational algorithm may be designed to determine whether slides produced by the laboratory or WSIs generated by a scanner are appropriate for evaluation and may even be designed to automatically trigger a recut or a rescan if found to be insufficient (Fig. 1).²⁶ An operational algorithm may also be used in a screening capacity, in which cases or slides flagged by an algorithm as suspicious are triaged for pathologist review (or in the case of a second read after the pathologist makes their initial diagnosis, presented to them for re-review). Assistive algorithms, on the other hand, generally support a pathologist at the time of slide review, directing pathologists to potentially important regions of the slide, enhancing the slide viewing experience, or providing adjunct information not otherwise available to the pathologist during signout.

When operationalizing clinical AI models in pathology, there are many logistical issues that must be undertaken. For example, AI requires data sources to be integrated and fed to the model, with new data pipelines developed. If additional clinical data, separate from WSI, is required, new interfaces will be needed in the electronic health record (EHR) and/or the laboratory information system (LIS). Once an AI model output is produced, a decision then needs to be made regarding whether that data should go back to a clinical health information system (HIS), be it the LIS, EHR, or pathology image management system. In some cases, the pathologist may work independently within an AI platform without HIS integration, whereas in other cases, this will be required so the pathologist can best act on and incorporate the AI model outputs into their clinical workflow. Integration processes are still in their infancy, with standards for AI data exchange lacking overall.

In addition, there is the question of whether AI models should be considered to be part of an existing lab test, a new discrete lab test, or otherwise separate from lab testing in general. Current Clinical Laboratory Improvement Amendments (CLIA) regulations stipulate a clinical laboratory be defined as a “facility for the biological, microbiological, serological, chemical, immuno-hematological, hematological, biophysical, cytological, pathological, or other examination of materials derived from the human body for the purpose of providing information for the diagnosis, prevention, or treatment of any disease or impairment of, or the assessment of the health of, human beings.”²⁷ From the authors’ own personal experiences, many debates have taken place over the past few years as to whether data derived from specimens should be considered to be “materials derived from the human body,” raising the question as to whether AI models used in pathology are subject to formal CLIA regulation. Until that matter is clearly settled within CLIA and other federal agencies, it is up to the performing laboratory implementing AI models to determine how best to validate any AI models used within their practice.

Finally, a single algorithm can have broad application supporting either operational or assistive applications, such as a screening, QC, or interactive real-time feedback. For example, grading a cancer based on histopathologic criteria often is subjective and in many cases can lead to either overgrading or undergrading of a patient’s tumor. AI-based cancer detection and grading tools therefore have the ability to be effective screening tools, interactive guides to assist the pathologist with interpretation,²⁸ or virtual second opinions.²⁹ In these situations, however, even though the algorithm is the same, each separate use case, otherwise known as the algorithm’s three different intended uses, must be validated. Importantly, the pathologist’s use of the model, rather than the model in isolation, should be evaluated, as pathologist behavior may also change in response to the introduction of these new tools.

3.3.

How Do We Optimize Clinical Utility Through Interoperability with Other Omics?

Although the prior discussion on the clinical utility of AI tools has been focused mostly on the field of pathology, the reality is that medicine is increasingly moving to multidisciplinary approaches that rely extensively on multiple diagnostic modalities. With the increasing complexity of clinical medicine, it is critical, now more than ever, to integrate the information content generated by the three major diagnostic modalities of clinical practice: radiology, pathology, and genomics. All of these diagnostic specialties can be considered “high-throughput” in terms of the raw patient data generated. For example, genomics produces a tremendous amount of sequencing data, such as deoxyribonucleic acid (DNA), ribonucleic acid (RNA), and epigenetic data. Radiology generates large amounts of imaging data, whereas pathology is the key repository for various forms of lab-based patient data, including both AP and laboratory medicine. In addition, in each specialty, there are extensive post-processing steps required to convert raw data into actionable diagnostic content impactful for both patient care and populating each of their respective omics.

Radiomics deals with the implementation of high-throughput, fully automated, computerized extractions of multiple quantitative image features, such as shapes, texture and pattern analysis, and intensity analysis from radiological images and data. These would be details that were part of the “experience” of a radiologist in the past. However, by explicitly quantifying these features upfront in an automated manner, the hope is to reduce inter-observer variation for more uniform outcomes. In addition, availability of this data lends to the mining of these quantitative features to aid in repeatable diagnosis and prognostic assessments.^30–32 Similarly, “pathomics” aims to extract, mine, and perform data analysis of subvisual features from a histology slide. Automated extraction and analysis of histopathological features using deep learning-based measurements can provide novel insights into the diagnosis and prognosis of patients that are pathologist experience independent. The goal is to use annotated WSIs to generate quantitative data, correlated with the histomorphological analysis in conjunction with the macroscopic radiological findings. Radiomics and pathomics deal with interpretable quantitative patient data generated based on morphological assessment algorithms, whereas genomics curates patient sample data at a molecular level. High-throughput sequencing data from a patient can provide information about the mutations, copy number variations, and other complex genetic alterations driving the disease at molecular level. Molecular pathway level information enables a finer spatial and temporal resolution of the progression of a patient’s disease pattern, providing novel means to diagnose as well as monitor the disease.

A current challenge in the use of diagnostic data is the siloed nature of patient data, i.e., using clinical data outside of the primary discipline that generates it. In most healthcare centers, the majority of patient data is essentially trapped in separate HIS with minimal cross-talk of content and/or expertise. In the case of AI, this is particularly important given that the more complex the algorithms become, such as with multimodal AI, the more we need to enable a clinically actionable omics mindset. Integration of patient metadata, clinical imaging (including pathology images), treatment plans, proposed interventions, and clinical outcomes is crucial to creating an enterprise data solution within a healthcare system. The foundation of such an end-to-end solution lies in the architecture of data, databases, information systems, and platforms that underpin the various medical specialties.

However, technology is not the only factor at play here—clinicians have a critical role in driving the interoperability required for such a solution. Interoperability implies the establishment and utilization of various standards of data in any given specialty. To start, clinicians can demand the use of standards in their everyday work and support efforts to integrate patient diagnostic data within the various clinical systems in use. They can further provide the expertise required to understand the various elements of patient data generated routinely in their discipline while concurrently working with vendors who can integrate patient data within the hardware, database, and application layers. Interoperability between major clinical systems, including the EHR, LIS, picture archiving and communication systems, data warehouses, and AI platforms, must be seamless with interfaces well established to ensure precise, secure, accurate, reliable, and uninterrupted exchange of relevant patient data points. A robustly designed layered information system architecture will thus allow an omics approach to AI, ideally with the different players (industry, hospitals/healthcare organizations, academic leaders, and providers/patients) driving the process forward. Finally, multiple AI approaches have been applied for multimodal analysis of WSIs in combination with genomics, transcriptomics, radiomics, and other large databases.³³ Thus, DP, in conjunction with an omics approach, can act as a key cornerstone to supplement the clinical utility of computational pathology and AI models within AP.

4.1.

Who is Needed to Enable AI in Anatomical Pathology?

Similar to work in the omics, much of the current effort in pathology AI is inherently multidisciplinary, requiring contributions from data scientists, pathology informaticists, administrative and operational staff, regulatory and clinical affairs experts, laboratory representatives, and pathologists. Each of these stakeholders brings a unique perspective to help shape the design of AI use cases and modifications to pathology workflows, as well as to provide familiarity with the costs, benefits, and potential pitfalls of implementation. The ability to understand limitations of a model, its susceptibility to preanalytical variability, or the limitations of the dataset on which it was trained or tested helps with predicting failures of a model in practice, which may not be immediately evident to each of the other stakeholders. It is important to solicit input from all pertinent stakeholders at the earliest stages of model assessment to avoid unforeseen challenges down the line, with particular engagement from the users who will likely be interacting with it the most; this may require including representation specific workgroups (such as practices with subspecialty signout) and early engagement of laboratory technicians who may be directly interacting with the model. Data scientists and engineers should be heavily involved in the model evaluation process to help practitioners interpret the outputs and to settle on solutions. Informaticists will usually work closely with stakeholders when developing workflow modifications, especially when there is an expectation that the model will interact with, or reside within, existing IT infrastructure (e.g., interfacing with the EHR, LIS, or other systems). Cybersecurity, compliance, data management, computing, and ongoing algorithm monitoring and maintenance are also important considerations; informaticists are well poised to contribute to these discussions and engage IT where appropriate.

Another set of relationships that are important to manage are those that exist outside the institution. As commercial offerings continue to emerge, this presents an opportunity for laboratories to implement solutions many times already deployed at multiple institutions, eliminating the need to develop algorithms from scratch (although not obviating the responsibility of validation and assessment). Of note, many institutions have a strict set of rules governing vendor relationships, starting with the procurement and risk management processes and extending through IT governance, infrastructure review, and data management. Even absent strict rules, it is essential to manage vendor engagement in a way that is responsible, ethical, and patient-centered. This means due diligence on the part of the practice is required at every step of deployment, including the following:

Finally, AI is very much at its earliest stages in pathology—it is burdensome for any single institution to be able to internally develop all of their potential AI needs. Therefore, vendor relationships are proving to be an important piece of the AI puzzle going forward, but overall they should be carefully managed, especially if co-development relationships are formed.³⁴

4.2.

DP Adoption Must Increase to provide the Substrate for AI

It is currently estimated that there are $\sim \mathrm{102,000}$ pathologists in practice today across $\sim 130$ countries worldwide. The bigger challenge is that the resources are not evenly distributed as a function of the population served. A recent analysis showed that two-thirds of the pathologist workforce is located in just 10 countries.³⁵ In countries where resource shortages are severe, such as Latin America and Africa, the use of AI tools has the potential to ease the pressures of these shortages by increasing diagnostic productivity and by providing elevated clinical decision support, thus increasing the quality of care. Importantly, AI tools must be introduced that are cost-effective, meet specific needs, and do not carry burdensome technical infrastructure requirements that cannot be met in low-resource settings. It is essential that the introduction of AI does not create greater health disparities, and therefore developing improved approaches to digitization, sharing, computation, as well as access to the relevant expertise, will represent an extremely important innovation.

There are many steps to take to encourage more widespread adoption, but some of these steps, including demonstration of clinical safety and acceptability for use, have already begun. For example, several peer-reviewed publications have shown concordance of primary digital slide diagnoses to traditional glass microscopy to demonstrate clinical safety,^36–38 and recent regulatory clearances and governing body guidelines show increasing support for DP in the clinical setting. For practices that wish to deploy AI, these have been important hurdles to clear to eliminate the need for parallel work streams that rely on the glass slide for diagnostics and the WSI for AI assessment. Ongoing work is needed to understand how DP and AI together can provide benefits not achieved by traditional best practices.