The hematoxylin and eosin (H&E) stain has long been an essential tool in the pathologist’s arsenal. Positively-charged hematoxylin and negatively-charged eosin bind to different substances within a tissue sample. This allows pathologists to clearly view different structures within that sample in various shades of blue, violet, and red. Once stained, the tissue sample can be used by a pathologist for medical diagnosis of a patient. This diagnosis will inform the type of treatment the patient will receive.
While H&E stains provide key insights into a patient’s diagnosis, they are not without their flaws. Because diagnoses from H&E stains rely on the human element of a pathologist viewing them through a microscope, different pathologists often come up with different diagnoses when viewing the same tissue sample. These pathologists are often extremely experienced, sometimes specializing on a single type of tissue, but there will always be some variation due to the subjectivity of viewing tissue through a microscope. To help mitigate this, pathologists often rely on algorithmic decision trees to group patients into more easily reproducible categories. This can help promote consistency between the diagnoses of different pathologists, but discrepancies still exist. Such decision trees may also run the risk of failing to account for more subtle details in a tissue sample.
The drawbacks of the H&E stain have caused some to turn to other methods for diagnosis, such as molecular testing. Molecular analyses can offer similar insights as the H&E stain in a much more data-driven, reproducible manner. However, such analyses can often be time-consuming or expensive compared to the H&E stain. Furthermore, the H&E stain retains other advantages over molecular testing. For example, H&E stains preserve potentially important spatial information about a patient’s tissue sample in ways that a molecular analysis cannot. H&E stains can also be used to make decisions during the course of a surgical procedure, for which a molecular analysis would be too slow.
What if there was a tool that could mitigate the downsides of the H&E stain, while retaining or even amplifying its benefits over other methods of diagnosis? Thanks to the recent rise of whole slide imaging technologies, artificial intelligence (AI) could be that tool. A type of AI known as a deep neural network (DNN) can be trained on sample whole-slide images to recognize different patterns and features within an H&E stained tissue sample. Once trained, this DNN can be fed new images and it will output reproducible, quantitative information about a patient’s diagnosis.
Current AI algorithms have already begun to demonstrate that they can diagnose certain medical conditions with high accuracy and low false-negative rates when compared to pathologists. Such algorithms will not replace pathologists, however. It’s more likely that AI will become an essential tool for pathologists, providing them with important information to be used in their diagnosis. Using AI as a tool may eventually allow general pathologists to provide specialist-level diagnosis for patients.
More benefits of AI analysis on H&E stains
- DNNs can identify subtle features that would be time-consuming or difficult for a pathologist to identify.
- Patients can receive more personalized treatment based on the minute details of their specific case.
- The marriage of AI and pathologists will speed up the diagnosis process considerably.
- AI will help reduce the pathologist’s work load, which will mitigate the shortage of pathologists that is expected to worsen in the future (2).
- Quality control and second opinions will be much less time-consuming.
The benefits of artificial intelligence in pathology are clear. AI can mitigate the downsides of the H&E stain while preserving the benefits when compared to other methods of diagnosis. The computational pathology team at Reveal Biosciences uses deep learning to analyze whole slide images of H&E stains, generating consistent, quantitative data. This data can be used to accelerate research and benefit patients.
Contact us to learn how we can apply these methods to your samples.
- Djuric, Ugljesa, et al. “Precision histology: how deep learning is poised to revitalize histomorphology for personalized cancer care.” npj Precision Oncology 1.1 (2017): 22.
- Robboy, Stanley J., et al. “Pathologist workforce in the United States: I. Development of a predictive model to examine factors influencing supply.” Archives of Pathology and Laboratory Medicine 137.12 (2013): 1723-1732.
Jeremy Warner, Scientific Marketing Associate, Reveal Biosciences