I am excited that AI tools are becoming widely available to extract information from experimental histopathology samples. In research settings, AI has many advantages over visual scoring and manual annotation – in particular, eliminating some tedious and time-consuming tasks; reducing inter-pathologist variability; and offering greater capacity to analyze archived samples. I’ve started to train, validate, test, and use my own algorithms to rapidly identify and quantify regions of interest within Mycobacterium tuberculosis-infected lungs. AI can speed up data acquisition – and, unlike human pathologists, algorithms happily operate 24/7. This means that I can set data-extraction and low-level analytical tasks to run overnight. AI cannot replace a pathologist’s interpretation and intellectual contribution regarding mechanisms of disease or hypothesis generation and testing. However, the two – human intelligence and artificial intelligence – can certainly complement each other.
After training and validation, algorithms can detect, quantify, and spatially locate tissue, cell, and subcellular features. In research settings, pathologists often score, grade, or quantify visual changes in tissues where the diagnosis is already known. Many research pathologists spend a substantial amount of time quantifying patterns, rather than establishing diagnoses. For example, in my own research laboratory we study host responses to M. tuberculosis. All samples come with a diagnosis: tuberculosis. Similar scenarios occur across many research fields each day. A major benefit of AI for research pathology is automatic recognition and quantification of visual information. Thus, AI algorithms can transform complex visual patterns into rich, quantified data sets that can be rigorously analyzed by statistical or machine learning methods. A second benefit is that AI doesn’t need sleep or caffeine.
Like in human medicine, the emerging use of AI in veterinary medicine is a disruptive technology that needs early adopters to gather data and feedback in testing phases, critical evaluation of the pros and cons, and a rational path forward. All aspects of digital pathology, including AI, are now being examined by our professional organization, the American College of Veterinary Pathologists. We are engaging in discussions across many sectors: industry, academia, and diagnostic and research laboratories. We have challenges to be addressed and overcome. Some of those challenges are physical or resource-related: access to equipment (scanners, servers, image-sharing software); hiring additional staff to perform and quality-control scans; and more. Other barriers are within our minds: fear of the unknown; fear of losing jobs; fear of becoming obsolete. These are scary concepts that have no easy solutions.
In our profession, we have realized that faculty who are training veterinary pathology residents need reliable information on digital pathology and AI. They need equipment and time to train themselves to become comfortable using it. And they need to understand the limits of this new tool. What I find interesting and exciting is that current trainees and recently boarded pathologists have found and adopted AI on their own through other resources. Some have produced and tested algorithms of their own; others are skilled coders who know more than their teachers (which I think is a great thing). The future of digital and computational pathology is bright, and education is the key. We must educate both ourselves and the next generation.
Newsletters
Receive the latest pathology news, personalities, education, and career development – weekly to your inbox.
