Artificial Intelligence for Computational and Digital Pathology
Part Two: Key Elements and Applications
In this second installment of our two-part series, Eric Walk, M.D., chief medical officer at PathAI, discusses the application of artificial intelligence in pathology. This Q&A session with Roche Diagnostics explores the different types of AI, its specific impact on the field and what pathologists can expect from the technology in the future.
What is the future of lab pathology?
Agentic AI Agents, capable of understanding both visual data from digitized slides and associated textual information, will enable sophisticated pathology decision-support systems. These systems could automatically analyze images, suggest diagnoses, propose relevant ancillary tests and even draft pathology reports. Pathologists will then function more as reviewers and approvers, leveraging AI insights to enhance accuracy, efficiency and consistency. Ultimately, this evolution promises to reduce information overload, improve diagnostic precision, and facilitate better, more timely patient care, potentially changing the pathologist's role to that of a strategic architect of the patient’s diagnostics workup, leveraging digital and AI tools.
How does AI work within a digital pathology workflow?
In a pathology workflow, AI analyzes digitized images of tissue samples, which are created by scanning traditional glass slides. Machine learning algorithms, particularly deep-learning neural networks, are trained on large datasets of annotated images, in which expert pathologists have identified and labeled specific features, such as cancerous cells or tissue structures. Once trained, the AI can automatically detect and quantify these features in new images, assisting pathologists in tasks like tumor detection, diagnosis and grading, and biomarker analysis. This helps to increase efficiency, accuracy and reproducibility in the diagnostic process, and can also aid in prioritizing cases and predicting molecular alterations from routine hematoxylin and eosin stains.
Why is it important that we incorporate AI into the pathology workflow?
AI pathology has the potential to solve two major challenges facing the field of pathology, increasing demands on workflow efficiency and information overload.
The vast majority of labs I speak with describe a need for greater workflow efficiency. Due to laboratory consolidation, organic growth in disease prevalence and other factors, many labs have increasing sample accessions and the same or fewer laboratory and pathologist resources. This includes AI-enabled slide quality control, case distribution, reflex test ordering, pathologist case prioritization and assisted reporting, all of which can be thought of as AI-based decision support for histotechnicians, lab directors and pathologists.
The second major issue facing pathology is information overload. Pathologists today are required to be up-to-date and knowledgeable about a large and increasing number of information categories related to patient diagnosis and care. AI technology, specifically large language and visual language models and chatbots, has the potential to serve pathologists by integrating these vast information streams into actionable knowledge relevant to the case and patient at hand. In this way, AI-powered pathologist decision support becomes a means for more accurate, informative and efficient reporting to both clinicians and patients.
Digital vs. Computational/AI Pathology
Pathology
- Study of causes, nature, and effects of disease
- Examine tissues, organs, and body fluids
Computational Pathology
- Combines AI and DP to extract meaningful information
- Develop algorithms to automate pathology tasks and processes
Digital Pathology
- Digitization of pathology specimen (WSI)
- Electronic analysis and sharing of data and reports
Digital Computational Pathology Tasks
- Tumor detection, classification, localization, segmentation
- IHC, ISH grading, scoring
- Morphological subtyping and feature analysis
- Disease diagnosis, quantification, classification, clustering
- Cell analysis (nuclear pleomorphism, cell crowding, cell polarity, mitosis, color, threshold, categorical threshold, object size, heterogeneity)
- Mutation status and burden
- Biomarker discovery (nuclear, cytoplasmic, membranes)
- Defective DNA
- Therapy response
- Survival forecast
- Clinical outcomer prediction
- ROI annotation (host tissue, target tissue, blood vessels, stroma, tumor, organ compartments, etc.)
- Cell counting
- Stromal feature extraction/morphometry
- Rare event screening (highlighting samples, micrometastases)
- Next-generation morphology (extracting new patterns from digital images, clinical correlations)
- Automated management and prioritization of pathology workflows
- Digital image analysis (color correction, filtering, edge detection, pixel intensity thresholding, mathematical transformations)
- Quality assurance/quality control
Can you compare today’s histology workflow to what it may look like in the future when AI is more integrated?
In the future, with increased AI integration, the workflow will be significantly streamlined and augmented. Digital pathology will be the norm, with slides being scanned and digitized. AI algorithms will pre-analyze these digital images, automatically identifying regions of interest, quantifying features like tumor cells or specific biomarkers, and providing preliminary diagnoses or scores. Pathologists will then review the AI's findings, focusing on complex cases and making final decisions with the aid of detailed, objective data, leading to faster, more accurate and consistent results.
One of the more common pathology uses of AI is in AI-assisted scoring of tumor-cell fraction. Can you explain how that works?
Tumor-cell fraction or tumor-sufficiency assessment, is a standard workflow step prior to molecular/next-generation sequencing testing. Today, this task is done manually by human pathologists, taking valuable time away from their primary diagnostic work. AI-assisted scoring of tumor-cell fraction involves using machine-learning algorithms to analyze digital images of tissue samples. The AI is trained on a large dataset of images where expert pathologists have annotated and identified tumor cells.
Once trained, the AI can automatically analyze new images, accurately identifying and outlining areas containing tumor cells. It then calculates the percentage of the tissue sample that is made up of tumor cells, providing a precise, quantitative score of the tumor-cell fraction. This process assists pathologists by offering faster, more consistent and less subjective assessments compared to manual estimation, aiding in diagnosis, treatment planning and molecular testing.
There’s been some interest around a non-oncology disease, MASH, liver disease. Can you give us an overview?
In non-oncology diseases like MASH, known also as metabolic dysfunction-associated steatohepatitis, AI works by analyzing digital images of liver biopsies to identify and quantify key cellular, inflammatory and fibrosis-related features used to grade MASH and critical for MASH clinical trials. These features include macrovesicular steatosis, lobular inflammation, ballooning degeneration and fibrosis. AI algorithms, often trained on datasets of annotated images, learn to recognize patterns associated with different stages of MASH.
By automating the analysis and providing precise measurements of these features, AI can improve the reproducibility and accuracy of disease scoring, potentially overcoming the subjectivity and variability inherent in traditional manual assessment. This can lead to more reliable diagnosis, better stratification of patients in clinical trials, and ultimately, more effective treatments for liver disease.
To understand more broadly the different types of AI, view the first part of our series.
Contributor
Eric Walk, M.D., FCAP, is chief medical officer at PathAI, where he guides precision medicine and regulatory strategy. He has more than 20 years of precision medicine experience, including overseeing FDA clearance and approval for multiple companion diagnostics and digital pathology algorithms at Roche Tissue Diagnostics and Ventana Medical Systems.