Image Reliability is the Foundation of Computational Pathology

I participated in a global panel discussion to examine the digital transformation of pathology labs and practices, along with Dr. Cory Roberts and Dr. Rajesh Dash. We talked about progress in establishing standards, emerging best practices related to calculating return on investment, and experiences with digital implementation.

At Seoul National University College of Medicine, we began a journey to digital in 2018 and now practice digital and computational pathology. Although pathologists, histotechnicians, and support staff readily accept digital workflows, the pace of adaptation to digital pathology is slow. The speed of transition to on-screen diagnosis and gaining confidence in image quality and image reliability seems dependent on a number of factors related to the monitor and gaining confidence in image quality and image reliability.

The first factor is pathologists’ confidence in making a diagnosis on a monitor. On-screen diagnosis introduces new considerations that impact pathologists’ daily practice. For example, the monitor screen has a larger field of view than a microscope, and therefore, it takes additional time to view a whole slide. H&E colors may display differently via a monitor, which may require an adjustment period for the user. On-screen diagnosis also requires changes in practice patterns for histotechnicians, who must ensure that all slides are free from mechanical elements that could cause malfunction of scanners or autofocusing errors.

A second factor is a need for discussion within the pathology community on quality control of digital images, and if and when to employ advanced technologies to verify the quality or reliability of images. This need provides an opportunity for pathology teams to apply new techniques to assess the quality of staining and provide feedback to laboratories for quality control. At my institution, we have experienced the common causes of errors, such as dust or air bubbles in core biopsies. Collecting baseline data to assess the range in which the analysis algorithm or autofocus functions worked well or not can be used to develop an automatic technique to detect errors and overcome them in the future.

I see a paradox in that pathologists and histotechnicians are spending more time and effort in order to use automatic or digitizing technology in pathology. In radiologic imaging, there have been many attempts to use AI technology to improve image quality and try to verify them by non-inferiority testing compared to traditional imaging protocols. Although there are many objections to using synthetic or artificially manipulated images for a confirmatory diagnosis like pathology, I think that reliability can be sufficiently obtained through a comparative test with a conventional system such as a digital pathology system which has been approved for clinical use.

Looking forward, we are starting to see that deploying AI technology on pathology images and molecular data can drive more quantitative and analytic biology and additional histologic information from patients’ samples. Many in vivo diagnostic skills can be developed to select patients requiring additional diagnostic tests, so they don’t lose treatment opportunities or undergo unnecessary biopsies. For example, the integration of radiomics and special information of histologic features is a good challenge to use for in vivo diagnosis technology and histology. I am intrigued by the potential presented by slide-free images and support the evolution of traditional H&E slides to digital images.

As our profession undertakes the transition from traditional methods to digital and computational pathology, there is much to learn from one another. Resources, such as The Leeds Guide to Digital Pathology, are helpful models for sharing real-world experiences. We can learn together from daily experience to accelerate our collective transition to a digital future.