By applying artificial intelligence (AI) tools to detect patterns in cellular and radiology images, a nationally respected data scientist is advancing clinicians’ ability to deliver precision medicine to their cancer patients.

“AI can add value in many ways, such as determining the presence of cancer, whether or not the cells are likely to spread, and how well a tumor will respond to treatment,” said Anant Madabhushi, Ph.D., Donnell Institute Professor of Biomedical Engineering at Case Western Reserve University’s School of Engineering and founding director of its Center for Computational Imaging and Personalized Diagnostics (CCIPD). “AI can also play a role in preventing cancers, such as assessing the risk in patients with other chronic conditions.”

Dr. Madabhushi spoke at a May 4 webinar sponsored by the University of Miami Institute for Data Science and Computing (IDSC). His talk on “Artificial Intelligence Across Scales: Implications for Precision Medicine” was part of the Data Citizens Distinguished Lecture Series, co-sponsored by the Miami Clinical and Translational Science Institute. He was introduced by Kenneth W. Goodman, Ph.D., director of data ethics and society for the UM Institute for Data Science & Computing (IDSC); founder and director of the University of Miami Miller School of Medicine Institute for Bioethics and Health Policy; and co-director of the university’s Ethics Programs.

 

Applying AI to Cancer

In his talk, Dr. Madabhushi focused on how AI can be applied to predicting disease outcome, recurrence, progression, and response to therapy in breast, prostate, brain, rectal, oropharyngeal, and lung cancers.

Noting that cancer will affect one in two men and one in three women during their lifetimes, Dr. Madabhushi emphasized early detection and appropriate treatment are the keys to saving lives. However, some interventions cause toxicity that may outweigh the treatment benefits—particularly for breast and prostate cancers—while new treatments such as immunotherapies are extremely expensive for patients.

“There are subtle patterns both in and around a tumor that can be detected through machine learning and AI.”

Current assessment techniques have their limitations, said Dr. Madabhushi. For example, breast tumors have different types of cancer cells, so a tissue sampling for molecular profiling could lead to different risk assessments depending the location of the biopsy.

“Breast cancer incidence also differs across populations, and we want to develop AI models that are specific to those groups,” Dr. Madabhushi said. His work has already demonstrated that an AI model specific to South Asian women delivered more accurate results than a population-agnostic model.

In general, AI tools that combine radiology images with pathology slides can provide better overall assessments than looking at just one dataset, Dr. Madabhushi explained. “There are subtle patterns both in and around a tumor that can be detected through machine learning and AI,” he said. “This is a low-cost approach with global impact, as these models could be accessed by clinicians anywhere in the world to provide a risk score for their patients.”

 

Assessing Risks and Outcomes

Taking digitized pathology slides from women with breast cancer, Dr. Madabhushi created high-resolution images to visualize the cellular complexity of a tumor. He then used deep-learning tools for identifying targets of interest to develop an Image-based Risk Score (IbRIS).

“With deep learning, we can identify features relevant to the biology of the disease and associate them with outcomes and treatment response,” he said. One of those features is the arrangement of collagen fibers in the cell. “Collagen is a highway on which cancer cells hitch a ride,” he added. “The smoother the highway, the easier it is for them to move to distant locations. Cancer patients whose collagen fibers were disorganized had better outcomes than patients whose fibers were highly structured.”

“AI assessment tools have the potential to advance precision medicine in other fields besides cancer.”

These AI approaches can be applied to other cellular features, as well as the microenvironment surrounding a tumor, Dr. Madabhushi said. He noted that a twisted network of blood vessels to a tumor is associated with a poor therapeutic response since this may reduce the delivery of chemotherapy medications or immunotherapies.

In prostate cancer patients, Dr. Madabhushi found that applying AI to images of glan-based features led to better predictions of post-surgical recurrence. He and his team have also studied lung, gynecological, and head and neck cancers, as well as brain tumors, and other forms of the disease.

Looking ahead, Dr. Madabhushi said, AI assessment tools have the potential to advance precision medicine in other fields besides cancer. “By looking at the architecture and interplay of cell types in the heart, we may be able to make better predictions of cardiac rejection rates in heart patients,” he said. “These tools also have important implications for treating patients with diabetes or other long-term chronic comorbidities.”

 

Story by Richard Westlund