In 2016, we at Sectra implemented our first image-based AI application in routine diagnostics at several labs. It was a computer aided detection tool, a KI67 cell counter that classifies positive and negative cell nuclei and presents the results to the pathologist. What was special was the unique visualization that allows pathologists to easily modify the classification before verifying the result. It has made it a popular tool among our users, and most of our customers who use it have decided to make it available to all pathologists in the lab. In addition to our own image analysis application, we are now offering integration with numerous other image-based AI applications, both CE-labeled commercial applications and research applications in routine diagnostics.
We have now reached a point where we have both knowledge and experience in implementing image-based AI in clinical practice, and we will be happy to share our experiences with the pathology community.
In pathology, the phrase “artificial intelligence” (AI) often brings image analysis to mind, even though AI can also be used for many different purposes, such as speech recognition and other applications. This article addresses how image analysis, what we call image-based AI, is used in clinical practice.
Normally, the objectives in using a specific AI tool are to increase quality through greater accuracy or boosting productivity through enhanced efficiency or a combination of the two. The need for enhanced efficiency is a consequence of the increased workload in the pathology community due to the increasing number of cases in combination with more complex cases. This calls for image-based AI tools that increase efficiency to boost productivity without lengthening the workday.
With a focus on quality, AI applications address the variability in pathology diagnostics. The individual interpretation of the case leads to variability between individual pathologists and laboratories and may impact both diagnosis and suggestions for treatment. AI applications addressing quality improvements is aiming to support pathologists when reading cases in order to minimize this variability.
Quantification is one area of application for AI; the aforementioned KI67 cell counter is one example of an image-based AI tool being used for that purpose. Scoring (or grading) is another use case. It is similar to quantification but is intended to classify tissue into categories on a continuous scale and not binary counting. Both quantification and scoring have the potential to improve accuracy in diagnostics while decreasing variability among individuals and laboratories. Of course, some of these tools will also save time since counting hundreds or thousands of cells is a tedious and time-consuming task.
Image-based AI can also be used to speed up reviews by improving the workflow for the pathologist. One example that Sectra has implemented is automatic alignment of different sections from the same block so that H&E and IHC stainings can be compared quickly side by side using synchronized panning and zooming. Other examples of image-based AI efficiency enhancements include assisting the pathologist in getting started on a case or glass with the highest probability of a relevant finding, to save time on searches.
For each organ type and disease there is potential for implementing multiple image-based AI tools to improve both quality and efficiency, which requires access to a broad portfolio of applications that must be tightly integrated into the workflow.
The uses of image-based AI vary between applications and use cases. One factor to consider is how it is initiated: on demand by the pathologist or as pre-processing before reaching the pathologist. On-demand tools are initiated by the pathologist from a toolbox when needed for a selected use case, while pre-processing is applied to all cases or slides in accordance with selected criteria for a specific workflow. The results can be presented to the pathologist, either directly in the case overview or when the case is opened. Pre-processed analyses can also be used to orchestrate the workflow, i.e. prioritizing a case that the AI application has identified as critical.
The interaction with the application is very important and often an underestimated aspect. It will vary depending on if the AI application presents the pathologist with an unmodifiable result or a modifiable suggestion. Depending on use case, the different methods have different advantages.
There are clearly many use cases for image-based AI in pathology. The ability of the pathology image management system to handle integration with the image-based AI tools is critical for successfully implementing AI applications in a clinical setting. Users should be able to access the tools without needing to log in to different systems, and the AI results should be stored and accessible within the image management system for a consolidated patient view. Depending on the results, worklists can be populated, and cases prioritized to orchestrate the workflow accordingly. Only then can image-based AI be part of the daily workflow.
An image management system must be designed to handle the large number of integrations needed with open interfaces (APIs) and the use of image standards. Sectra’s digital pathology solutions have this integration potential; users can run either certified or research applications depending on need. This applies not only to applications that are available now, but also to future developments in the field.
Standards throughout the workflow, from scanner to pathologist, are vital for ensuring interoperability. Scanning images in DICOM ensure that images can be read and interpreted by the desired applications now and in the future.
At the time of writing, there are CE-labeled image-based AI applications from five vendors in the global pathology market. Moreover, there are proprietary image analysis from scanners vendors, research products and home-built applications. My prediction is that more analyses will come, from both existing vendors and from new ones. Self-validation is a time-consuming task that requires a great deal of energy, dedication, and knowledge, which some institutions with in-house capabilities can manage. For most laboratories, however, selecting applications that have put in the required effort to ensure performance and quality and to guarantee future maintenance is crucial. The number of image-based AI applications that will be required has yet to be determined, but it is certain to be large. Use cases vary depending on different organ types, stainings, diseases and treatments and so forth. Selecting an open-format image management system may help you successfully implement all the desired image-based AI applications.
To reap the full benefits of digital pathology and adopt AI in the most efficient way, having the digital pathology infrastructure in place with a digital pathology image management system that is AI-ready and capable of digital workflow management is crucial. The image management system needs to be well integrated with scanners and LIS to enable such a workflow.
Once the infrastructure is in place and everyday reading has been digitized, you will be well equipped to take the next step. Only then will your lab leverage all the new technology and be able to deliver greater accuracy and efficiency, which will ultimately both improve the work situation for many pathologists and be of benefit for patient care.
