Every day, hospitals generate millions of pathology images used to diagnose cancer. While standard Hematoxylin and Eosin (H&E) slides reveal structural details about tumor tissue, important genetic information often remains hidden.
A study led by Lana Garmire, Ph.D., professor and vice chair of research in the Department of Biomedical Informatics and Data Science, was published in Nature Biomedical Engineering. The study introduces HisToSpatialCNV, an artificial intelligence tool that uses a graph neural network model to predict spatial Copy Number Variations (CNVs) directly from routine H&E pathology images, for the first time.
CNVs are DNA changes that give cells extra or missing copies of genes, which can alter how cells function and contribute to diseases such as cancer. While current AI methods can predict some molecular features from pathology images, accurately inferring CNVs at high spatial resolution remains a challenge. Technologies that directly measure spatial CNVs also remain limited and are not widely accessible.
"This study addresses two major gaps in both digital pathology and spatial genomics," Garmire said. "Building on advances in graph neural networks and AI, we developed HisToSpatialCNV."
The project brings together expertise in artificial intelligence, computational biology, spatial genomics, pathology, and oncology. Collaboration across the Department of Biomedical Informatics and Data Science, the O'Neal Comprehensive Cancer Center and the Department of Pathology will provide access to clinical and pathology resources needed to further validate the technology.
Garmire says the interdisciplinary approach is important for translating the technology into patient care. The team combines expertise in advanced AI models and spatial genomics while working closely with clinicians "to solve problems that are important for patient care."
HisToSpatialCNV could help researchers and clinicians better understand tumor evolution, predict patient outcomes such as survival, and identify patients who may show different treatment responses. The approach could also provide important spatial genomic information directly from routine pathology slides.
Garmire hopes the technology will ultimately make precision oncology more accessible.
"When a patient undergoes a routine tumor biopsy, we envision being able to infer important spatial genomic alterations directly from the pathology slides," she said. This information could help clinicians "select more personalized treatment strategies" without requiring additional costly and time-consuming molecular tests.
The broader goal is to uncover genetic information already present in routinely collected pathology images and expand access to precision cancer medicine.
"We are teaching AI to "read between the lines" of these images," Garmire said. "Our goal is to make precision cancer medicine more affordable, more accessible, and available to many more patients here at UAB and around the world."