The method by which a piece of tissue becomes a diagnosis is not frequently known to those outside of the practice of pathology, and it is prudent to understand the details of how glass slides are made in order to understand the potential preanalytical variance and how that relates to training AI on pathology data; it can also give some insight as to how the practice of oral pathology is changing (and may change) with the application of AI.

Once a piece of tissue comes into the laboratory (often stored in a fixative like formalin, which irreversibly preserves the tissue and increases its firmness), it goes through multiple steps before it comes a glass slide ready for review ( Box 1 ). First, it is accessioned into the laboratory information system so it can be electronically tracked through every step of the process. The tissue is then “grossed,” where it is examined with the naked eye, described in the report, and cut into smaller pieces if it is a large piece of tissue. The tissue is placed into small boxes (“cassettes”), which are then filled with paraffin wax; these are called “blocks.” The blocks are then taken to a microtome (functionally a medical-grade, highly precise deli slicer) to cut the wax and tissue into thin (5–15 μm), nearly transparent slices, which are placed onto glass slides. These must be stained before review; almost all tissues relevant to oral pathology are stained with hematoxylin and eosin (H&E) by default. Following staining, a coverslip is placed over the tissue. This is then reviewed by a pathologist, who looks at the slide under a microscope and provides a diagnosis by writing a pathology report, usually by typing semistructured free text. The report is then released in a “final” state; further comments to a report may be provided in an “addendum,” or the report may be modified in an “amendment.” In more diagnostically complex cases, the pathologist may request additional stains to evaluate the presence of certain cellular staining patterns that may assist in providing a more specific diagnosis or provide predictive or prognostic variables relevant to the patient’s disease.

To enable the use of AI on histologic tissue sections, the slide must be digitized, ideally at high magnification. To do so, significant costs are incurred: an entire digital pathology workflow for clinical use requires purchasing slide scanners, digital storage space, a digital slide viewing software, high-resolution monitors for viewing the slides, and maintenance for all those components. These costs, unlike the transition from handwritten clinical notes to electronic health records and film radiology to digital radiology, are additive rather than a replacement for the old method—digitization of slides involve an additional step following the conventional workflow. Laboratory leadership must also consider modifying their existing workflow to optimize the slide scanning process (eg, at our institution, we chose to switch to film coverslips to avoid errors caused by misaligned glass coverslips and excess adhesive). Institutions that have adopted digital pathology workflows have done so to centralize processing while allowing pathologists to work remotely, pursue cost savings relating to storage of physical media, and to prepare for AI-enabled workflows.

The amount of data available depends on how much tissue is on the slide, the magnification used, and if multiple focal planes, or “z-stacks,” are scanned (in practice, z-stacking is not often used in oral pathology images). To provide an example, Table 1 contains metadata from 2 digital pathology images, 1 small biopsy and 1 large resection, that were originally scanned at 40× on the Philips Ultrafast Scanner (Koninklijke Philips N.V., Amsterdam, Netherlands) as ∗.isyntax files. Each was converted to TIFF files (with quality parameter set to 80, the default), exported once at 20× magnification and again at 40× magnification, and metadata was extracted using QuPath (an open-source digital pathology software). The TIFF file for the biopsy is significantly smaller than the resection, and at 20×, the files are smaller than at 40× because a digital slide scanned at 40× objective has 4 times the area and approximately 4 times as many pixels as a slide scanned with at 20×. While most oral pathology biopsy cases (in the author’s experience) consist of a single slide, a resection for head and neck cancer frequently consists of dozens of slides from multiple specimens: separately submitted margins, the resection itself, and lymph nodes. The specifics of the case can cause the amount of data generated to vary wildly; a single large resection case at our institution may use hundreds of gigabytes of storage.

The quantity of data has implications for considering the cost of implementing a digital pathology workflow and also for how AI works in these cases. One consideration would be to use lower magnifications for AI tasks, which would reduce the amount of data that needs to be processed. However, depending on the AI task, the magnification could make a difference in performance. Some sample images from the biopsy and resection earlier are provided in Fig. 1 . While little visual difference can be appreciated for the files exported at 20× and 40× resolution at low power (4×), as magnification increases the 20× file becomes blurrier compared to the 40× file. A task that relies on detailed nuclear features might benefit from the higher resolution 40× (and possibly a higher TIFF quality). However, tasks that rely on low-power features (eg, identifying large metastatic foci in lymph nodes) may be able to function reasonably at 20× or lower magnification. Regardless, for AI tasks in digital pathology, the resolution of the image (pixels per micrometer) is important, and should be reported in publications or when considering the implementation of an AI tool in pathology practice.

Visual comparison of digitized pathology for a biopsy and a resection, scanned at 20× and 40× magnification ( top ). Machine learning tasks can take images from a tissue level specified ( left ); while minimal visual differences can be seen at lower magnification, at higher magnification tissue features begin to lose definition, which can impact downstream performance even if the differences seem minimal to the human eye.

There are multiple implications that this workflow has on the use of AI in digital pathology, from a practical perspective and from a reporting/research perspective. Due to the size of the images, a single digital slide cannot be used in toto in an AI algorithm—most AI algorithms expect an image no larger than perhaps 1024 × 1024 pixels large, so researchers must have mechanisms to break the slide down into smaller components (regions of interest, ROI) that can be reviewed by AI. Combined with the highly variable amount of data available for each case, algorithms for research and for clinical use must be able to handle these data by either identifying areas of interest or performing some sort of aggregation technique to designate a single class for multiple regions and multiple slides. For example, a task where an AI is performing the pathologist’s primary task—providing a diagnosis on a single specimen—the AI must be able to review all ROIs on each slide, then consolidate the results from all slides for that specimen to provide a final diagnosis.

Another substantial issue that the overall pathology workflow imposes on the use of AI in digital pathology is that many steps in the slide preparation process are inconsistent among laboratories—sometimes, even within a single laboratory. Variability may be present in

• •

  Tissue sampling procedures (in resections especially, as sampling decisions are not necessarily consistent)

• •

  Artifactual changes, including changes from frozen sectioning before the slide is submitted for permanent/formalin-fixed sections

• •

  Quality of block cutting (variable tissue thickness and artifactual holes in the tissue)

• •

  H&E staining (particularly in the intensity of hematoxylin)

• •

  Immunohistochemical (IHC) staining (intensity of staining, background “blush”)

In addition to the prescanning differences, digital pathology may also be different among laboratories. There are multiple digital pathology vendors, and because there is not yet a consistent digital slide format, digital pathology scanners output a wide variety of different file types that each store image data differently. A further complication is that many of these file types are proprietary and full specifications are not always available for review. An AI model that is trained on one type of scanner may not translate well to a different scanner.

Due to the substantial preanalytical variance present in digital pathology data, extreme caution should be taken before using an “out-of-the-box” solution in a new setting. While there are mechanisms to address variability of input data (such as stain normalization ), AI models in digital pathology should undergo robust validation testing on real-world data in an institution before going “live,” and companies creating these models would ideally release some level of demographic information describing their dataset so health care professions can make educated decisions about the applicability of those models to their specific contexts.