Success Through Centralization?

Digital pathology’s benefits have been widely discussed – from alleviating staffing shortages to accelerating turnaround times; from increased consistency and accuracy to innovative teaching tools. But, in practice, those benefits have proven elusive. Why?

At its most intensive implementation, digital pathology has the potential to support primary clinical diagnosis. At a less intensive level, it can be used in targeted precision medicine applications, particularly in the area of immunohistochemistry tissue diagnostics. In all cases, though, the burdens of digitizing often outweigh the benefits – mainly because responsibility for the system’s performance rests with the local laboratory. Though many companies have created effective tools for managing slide images and streamlining workflow, they have done little to relieve the most time-intensive burden: the creation, validation, and maintenance of image analysis tools and diagnostic algorithms.

To comply with Clinical Laboratory Improvement Amendments (CLIA) regulations regarding laboratory-developed tests (LDTs), an artificial intelligence (AI) diagnostic algorithm must be thoroughly tested and maintained. Quality assurance procedures must be created and users provided with regular, comprehensive training. Digital pathology equipment must be purchased and the informatics expertise needed to integrate it into laboratory and hospital information systems must be established. The overhead necessary to support these activities is prohibitive for most laboratories in the current reimbursement environment. A potential solution: AI algorithms offered as LDTs on a national scale. To take advantage of this capability, clinical laboratories would send slides or digital images to a service provider and receive a clinically appropriate diagnostic result. This would leverage current digital pathology product strengths while mitigating the challenges associated with DIY solutions.

For instance, local laboratories have limited ability to design and execute a clinical study of sufficient power to verify and validate a lab-developed algorithm in line with CLIA requirements. Based on patient volumes at a typical hospital or regional laboratory, the investment of time and money needed would be prohibitively expensive. Leveraging national economies of scale, a central service provider could develop an algorithm and test it in a clinically robust study that could be published in leading peer-review journals available to the entire medical community. This approach would not only yield a clinically superior product, but also eliminate the “black box” nature of DIY algorithms. It would also eliminate the organizational and financial challenges associated with DIY digital pathology solutions. Creating and maintaining image-based AI solutions is costly, time-intensive, and requires informatics expertise not every organization has. Testing them requires expertise in study design and access to clinical samples, which may be difficult or expensive to procure. And, as digital imaging equipment and staining technologies evolve, algorithms must be re-evaluated and updated to keep pace. Quality assurance procedures must be created and maintained and employee training programs implemented. All of this would need to be achieved within the current clinical diagnostic reimbursement environment. Without additional funding, it is not surprising that few laboratories can generate the financial ROI to justify a DIY digital pathology approach.

A centralized digital pathology service offered on a national scale would offer crucial additional benefits. For instance, purchasing efficiency is critical – so this approach enables the creation of a family of solutions local laboratories can procure efficiently. Most laboratories prefer to consolidate their purchases through as few vendors as possible; a centralized digital pathology provider could offer a “one-stop shop.”

The provider would also be able to amass a powerful database that could be leveraged in support of clinical studies. A key characteristic of image analysis is that it generates large amounts of data. This data could support sophisticated clinical studies impossible to implement at the local level. Additionally, the data could be used in ways not otherwise possible, such as more robust evaluation of cut points and treatment decisions. For example, finding strong correlations between patient response to immunotherapy and currently available diagnostic data has been challenging. When studying such a complicated and dynamic clinical question, the larger the database the better. Imagine if a centralized provider – for instance, of PD-L1 AI solutions – stored the de-identified data from all of the samples it processed over time. Such a database would house an extensive range of clinical diagnostic data for tens of thousands of patients. When merged in a research setting with patient outcome data, it could yield true medical breakthroughs.

The final advantage to using a centralized send-out digital pathology solution is image storage and technical infrastructure. To achieve the degree of accuracy required for an AI solution to be practical across a wide range of samples, the image and resulting data files must be highly detailed, nuanced, and layered. This leads to not only large files that need to be carefully tracked, stored, and backed up for lengthy periods, but also millions of small files that must be effectively managed. This demands vast storage capacity and a highly redundant and available file system built around these specific needs. It also requires uniquely optimized, high-performance computational power to achieve the desired output in a viable amount of time. Any of these factors alone would be a major endeavor for a typical laboratory – but combining them into a cohesive, secure, compliant, and reliable ecosystem is an immense, complex undertaking. In my view, it is only by offering centralized solutions that we can get all labs on board with digital and computational pathology.

A local laboratory might be able to create an AI solution for a single indication, such as NSCLC, but this solution would not apply to other PD-L1 indications. It would probably not be wise for a laboratory to invest the time and money to develop a solution that is only applicable to a portion of its needs. Only a centralized service solution would deliver the economies of scale needed to support the creation of a family of PD-L1 solutions to meet the operational needs of local laboratories. And only a centralized service could promptly update its capabilities to keep up with the evolution of clinical care. Quickly adding new indications, stains, or digital scanners is only feasible for a large service solution provider.

In the immuno-oncology field, PD-L1 reactivity is only a portion of the clinical picture. Immunotherapy’s effectiveness depends on both the PD-L1 reactivity of the tumor cells and the immune microenvironment. Current tests provide only a PD-L1 result because it is relevant in selecting FDA-approved therapies and it is the only reimbursable result. The fact that additional information about the tumor microenvironment (TME) is not typically available does not mean that it is not clinically relevant; it only means that the additional IHC stains required to identify macrophages and lymphocytes in the TME are not reimbursed.

What if the TME could be assessed without any additional time or cost on the part of the clinical laboratory? Though not commercially realistic using manual review techniques, an advanced AI algorithm could achieve this goal. AI is far superior to the human eye in detecting patterns that allow cells to be identified and classified across an entire tissue sample. What if a digital pathology solution could not only accurately detect PD-L1 reactivity, but also count and locate macrophages and lymphocytes and assess levels of tissue necrosis? Would an oncologist treating challenging patients benefit from this information? Although medical decisions must be based on clinically rigorous study data, the migration of ideas from research to FDA-cleared treatment typically takes decades. Oncologists perpetually find themselves in a grey area between cutting-edge research and fully validated, FDA-cleared treatments. The pathologist’s goal is to provide treating physicians with the facts and tools they need to manage their patients. Artificial intelligence could be an important tool in fulfilling that mission – and only a centralized digital pathology provider would have the expertise, reach, and economies of scale to leverage that capability.