AI in Pathology: Transforming Oncology Diagnosis and Treatment
Expert Perspectives · Waqas Haque, MD, MPH · 2024-08-28
As oncologists, we’re constantly seeking ways to improve patient care and outcomes. In this month’s blog post, we will review recent developments in artificial intelligence (AI) for pathology to enhance our diagnostic capabilities and treatment decisions.
Dr. Eric Topol, Founder of Scripps Translational Research Institute, and Dr. Faisal Mahmood, Professor at Harvard Medical School, recently highlighted the pressing need for AI in pathology, given the challenges of inter-observer variability and difficulties in identifying cancers of unknown primary. AI systems, trained on vast datasets, can potentially reduce bias and improve detection rates, particularly for underserved populations who may present with advanced disease; however, the adoption of AI in pathology faces hurdles. Low digitization rates in many healthcare facilities and regulatory barriers for AI companies testing their software in clinical settings are significant obstacles. Despite these challenges, AI is poised to become a competitive advantage for forward-thinking hospitals in the coming years.
Figure 1. Predicting molecular alterations from H&E-stained whole-slide images
In resource-scarce medical systems, pathology funding is often overlooked, yet the digitalization of pathology opens doors for AI tools to improve efficiency and extract more information from tissue samples. A 2019 analysis from Memorial Sloan Kettering Cancer Center suggested that digital pathology implementation could lead to a 93% reduction in requests for physical slides. AI-assisted methods are showing promise in various applications, including human epidermal growth factor receptor (HER2) 1+ cases, programmed death-ligand 1 (PD-L1) scoring, and Ki-67 assessment.
To ensure the reliability of AI solutions, it’s crucial to implement proper software design principles and rigorous testing. The maturity of AI solutions can be evaluated using technology readiness levels (TRLs), and healthcare facilities must consider whether to adopt AI gradually or implement full-scale changes. For successful integration of AI in pathology, healthcare systems need robust information technology (IT) infrastructure, including broadband internet, sample tracking systems integrated with laboratory information systems (LIS), whole slide imaging (WSI) scanners, and appropriate workstations for pathologists.
Figure 2. Technology readiness levels in pathology
As we move towards equitable AI in oncology, it’s essential to address concerns about validation and applicability across diverse populations and settings. A Nature article on AI equity highlighted that the average validation set for FDA-approved AI tools between 2010 and 2020 included only 300 patients, which may be insufficient to ensure safety and efficacy. Moreover, most AI devices were assessed using only retrospective studies, with no high-risk devices receiving prospective evaluations.
We must also consider the global applicability of AI tools. For instance, an algorithm developed using a 3T magnetic resonance imaging (MRI) scanner might not be suitable for use in many African countries and other low- and middle-income countries (LMICs) where 1.5T scanners are more common.
As oncologists, it’s crucial that we stay informed about these developments in AI for pathology. By understanding both the potential and limitations of these technologies, we can advocate for their responsible implementation to improve cancer diagnosis and treatment for all patients, regardless of their geographic or socioeconomic status.