Machine Learning and Image Manipulation in Digital Pathology: The Case for Efficiency Enhancements
Much of the excitement related to digital pathology comes from the ability to rapidly move cases from one place to another. Remote access from home, speedy consultations, rapid review of archival material, and access to cases for clinical conferences are just a few of the benefits that accrue from moving from glass slide analog to a digital slide environment. All of these workflows can be accomplished with glass slides but require the presence of the physical case material. With digitization the need for the physical case is negated which leads to the obvious advantages that slides need not ever be “pulled” from the files, and multiple individuals can view the same case at the same time – think about slides traveling to conferences or for consultation which then are needed immediately for clinical care or quality assurance activities. Clearly the digital pathology laboratory facilitates these processes and makes them more efficient (and safe).
But there are any number of workflow examples that digitization enables that just cannot be done in the analog glass slide-based setting. These workflows represent the added benefits of digitization above and beyond what pathologists are becoming familiar with today in their routine practices. A wide variety of efficiency tools are made feasible by digitization. Viewer-based tools facilitate and make more accurate measurements and allow digital annotations for areas needed to be quickly located for conferences or education. Additional digital tools can also greatly increase efficiency. Image registration, or the spatial mating of serial sections, facilitates quick localization of areas or features needing to be analyzed using multiple studies, such as special stains or immunohistochemistry. In analog workflow, comparisons between serial sections require first locating the feature of interest on one slide, then on another slide. When this involves many sections, the process can be time consuming and tedious. With registered images, optimally displayed together in the viewer, the pathologist can co-localize those features, examine them together, and then move to additional features with each “slide” moving in unison with the others. Quantitative assessments, such as counting cells positive and negative for a specific immunohistochemical stain and then judging the intensity of the staining is yet another example of a digital “assist.” These efficiency applications are ONLY available in the digitized world.
Of course, the holy grail of digital applications is the development of machine learning-based algorithms which can diagnose and prognosticate diseases, separating them from benign and non-pathologic processes. Such applications will come eventually but will require substantial development, regulatory oversight, and cultural acceptance before becoming mainstream. A more gradual approach to artificial intelligence in digital pathology therefore might be more appropriate. As such, applications that work in a similar manner to the “efficiency” gains made by digital pathology replacing glass slide workflows, as noted above, might be a good place to start.
One major category of AI-based “efficiency” tools would be those designed to prescreen histopathology cases before they reach the pathologists’ viewing station. Examples include prescreening prostate core biopsies for areas having a high probability of representing prostatic adenocarcinoma. Quick localization of “high-risk” areas would allow pathologists to move quickly through cases and could even improve overall accuracy by avoiding false negative misses. The same prescreening evaluations could take place in other high-volume areas such as lymph node dissections, and in bladder, cervical, and colonic biopsies. In each of these examples, the algorithms are not “diagnostic” – the pathologist still evaluates each high-risk area and makes the final decision, but finding the positive cases is more efficient. Another efficiency algorithm could be localization of margins in tumor cases. The ability of an algorithm to find ink of a color used to mark a specific geographic margin and indicating which slide has which color would be a substantial assist to the pathologist.
One application that we have developed in prototype at Corista is the ability to accurately identify glomeruli in medical renal biopsies. We combined this AI-based finding algorithm with image registration of serial special stains to produce an efficient process that first finds a feature and then co-locates that feature in 4 stains simultaneously. The double efficiency of these 2 processes could improve the process of reviewing such biopsies from the standpoint of time and perhaps accuracy.1,2
Overall, digital pathology can improve existing glass slide workflows, but in addition, it will create additional and more efficient workflows that don’t even exist in the analog world. It will take time to move all of these initiatives into practice, but clearly having such advances available will drive more universal acceptance of this new modality. “Diagnostic” algorithms will come eventually. But “efficiency” algorithms will most likely lead the way.
- Wilbur DC, Pettus JR, Smith ML, Cornell LD, Andryushkin A, Wingard R, Wirch E. Using image registration and machine learning to develop a workstation tool for rapid analysis of glomeruli in medical renal biopsies. J Pathol Inform 2020;11:37.
- Wilbur DC, Smith ML, Cornell LD, Andryushkin A. Pettus JR. Identification of Glomeruli and Synchronized Review of Special Stains in Renal Biopsies by Machine Learning and Slide Registration: A Cross-Institutional Study. 2021 Apr 4. doi: 10.1111/his.14376. Epub ahead of print.
Source: Corista
