There has been a vast uptick in technological advancements in the laboratory industry for digital pathology and Artificial Intelligence solutions. In fact, AI is the buzzword of the day. What is the difference between algorithms, uses and accuracy?
It’s very easy to conflate the terminology and in some cases, the differences are very clear while in others the lines are a bit less distinguishable. In short, Artificial Intelligence (AI) and Machine Learning (ML) are closely related, but you can think of it more as a hierarchical relationship. They both relate to computers and programming; however, AI algorithms rely on human input and ML algorithms are taught. In other words, AI requires humans to provide inputs such as visual perception, interpretation, decision making, and results in algorithms that can make decisions based on data that they have been trained on.
Whereas ML algorithms are based on building models, or neural networks, that are trained to identify patterns and trends in information to make decisions. ML is used to classify data and recognize patterns; this is actually essential to many AI algorithms.
Image Analysis (IA) is more about extracting data for a purpose, or multiple purposes from an image. This is very useful in workflow application optimization and can provide significant added value to AI and ML, but more to digital pathology solutions and their adoption.
There are two factions of thought on the overall use of these technologies in patient diagnosis and care. The first is pro-innovative technology and this group sees the benefits of being able to use well-developed, trained and tested algorithms in a variety of workflows. Some of the professionals who fall in this community are for all uses, some are for a subset and some are for a limited set of uses. The second is anti-adoption.
Overall, these use cases fall into the following general categories:
Diagnosis
AI and ML algorithms are typically deployed for use in diagnosis. They can be used for preprocessing images to provide a pathologist with an interpretation of the specimen. This can include the image which typically overlays on the original and shows a ‘hot spot’ view of the cancer within that slide and can also provide textual and contextual data i.e.: percentage of proliferation, grade of cancer, number of positive and negative cells, etc. This type of algorithm can often be used as a region of interest (ROI) that is manually invoked by the pathologist on a slide or area of a slide that they are looking for more insight on. When integrated within a digital pathology solution, this can also provide a streamlined workflow to include capturing the image and output to include in a final report in support of the pathologist’s diagnosis. These algorithms can be used (if FDA cleared for diagnosis or validated as an LDT by the laboratory/pathologist) as diagnostic or as an aid to the pathologist in rendering their diagnosis.
Workflow Optimization
IA algorithms provide a streamlined workflow through the preprocessing analysis of images. These algorithms can perform such tasks as auto registration of images, ensuring that the images in a virtual slide tray are in the correct order, analyzing images to make a priority assessment and move cases with a higher probability of being positive, or more complex, higher in the pathologist’s worklist.
Algorithms also contribute to significant efficiency and accuracy gains in use by pathologists. They can perform calculations and computational work that software can do exponentially faster than a human. This also removes the need for tedious, manual tasks to be performed by high-quality resources, such as pathologists, and affords them the ability to focus on their expertise, the interpretation and diagnosis of the patient.
Quality Assurance or Control
Algorithms can also be configured, or deployed, to provide either an automatic review of all cases or be set to do a review of a predetermined number or percentage of cases to support regulatory requirements for review or for internal training and safety. In this use case, the algorithms would be run either before a case is signed out, but after it is diagnosed, or after it is signed out to flag any questionable diagnosis. It can highlight areas that may have been missed, flag a tumor that may have been identified as a grade 2 when it was determined to be a 4 by the algorithm, etc.
The second set of professionals in this community are not bought into the safety or use of algorithms in diagnostic practice. There is the argument that many of these are not fully tested, they are not able to be used “out of the box” for a variety of reasons leading to a wide variation in how images are viewed (color saturation, stain types, scanner file formats, etc.). Many of the algorithms are rated on their accuracy by a correlation factor and pathologists are not always in concordance with themselves 100% of the time, never mind with each other.
Then, there is the concern that drift results can occur. Recently, Fortune reported that a study conducted by Stanford University found that “Over just a few months, ChatGPT went from correctly answering a simple math problem 98% of the time to just 2%”. While this is not indicative of what a commercially available, tested pathology algorithm would encounter, it does add to the doubt of those who are in this community of adaptors.
In summary, algorithms, when tested and validated by each laboratory for deployment within a digital workflow solution can provide improved efficiency, streamlined workflow, reduce errors and increase the experience for the pathologist.
About the Author:
Lisa-Jean Clifford
COO & Chief Strategy Officer, Gestalt Diagnostics
For more than 2 decades, Lisa-Jean Clifford has been a noteworthy leader in the high-tech healthcare solutions space. Lisa-Jean’s passion for making a positive impact on the lives of patients through technology can be traced back to her tenure at McKesson and IDX, now GE Healthcare, where she served in vital business development and marketing roles and to Psyche Systems, an LIS solution provider, where she was the CEO for eleven years.
Now, recognized as an industry expert, she actively participates in numerous boards including the Association of Pathology Informatics and MLO’s Editorial Advisory Board. She is widely published in many top laboratory publications and noteworthy news sources, such as Forbes, CAP Today, Medical Laboratory Observer, and Health Data Management. Also, she is a highly sought-after speaker and focuses on delivering valuable content in critical areas such as lab automation including software and interoperability, digital pathology, AI in pathology, lab informatics, oncology, and women’s health.
Lisa-Jean’s success can be attributed to her perseverance, integrity, her high-regard for ethics, and her desire to continue to learn, grow, and move technology solutions in a forward direction for healthcare. Her collaboration with industry partners, customers, colleagues, and competitors combined with her commitment to exceptional customer relationships is what distinguishes her drive to foster a win-win for the healthcare industry as a whole.
