Through its decades-long practice of digitizing clinical images, radiology is providing an early proving ground for machine learning-based approaches for disease detection, diagnosis, and prognosis. With a vast, rich pool of radiological images, researchers have access to the raw material needed to develop and hone new disease-detection algorithms, which are helping to drive the AI revolution in health care. Now, other sub-specialties are beginning to follow suit.
For example, in the U.S., the ability to digitize pathology slides in a clinical setting recently became possible, with the first whole-slide imaging system for digital pathology approved by the FDA in April 2017. The size of digital pathology images far exceeds that of radiological images. Even though a slide contains just a small slice of human tissue, it is often viewed under a microscope at multiple magnifications, generating numerous fields of view — all of which must be digitally captured. In addition, radiological images are typically viewed in grayscale, whereas pathology often deals with full color images.
Despite these challenges, researchers in academia and biopharma are already beginning to harness new capabilities in digital pathology. For example, teams in the U.S., Ireland, and the Netherlands are creating automated, machine-learning-based methods to detect cancer cells on a digital pathology slide. Although the tools are still under development and not yet available for routine clinical use, the goal is to give pathologists smarter, less time-consuming, and more precise methods to determine which patients have cancer and which patients don’t — and to help pinpoint the most effective treatments. Initial work is aimed at detecting a handful of cancer types, including breast, prostate, lung, and colorectal cancers. Eventually, these AI-based diagnostics will achieve sufficient training to identify any solid tumors.
Investigators are also working to extend the power of digital pathology to disease prognosis. A research team based in California recently developed an automated method that detects nearly 10,000 different features within whole-slide pathology images to distinguish different lung cancer subtypes and predict patient survival. The group set out to explore ways of improving the treatment of two types of non-small cell lung cancer, adenocarcinoma and squamous cell carcinoma, which can vary dramatically in terms of the recommended paths for treatment and overall prognosis, and are often difficult to differentiate based on microscopic inspection alone.
To tackle this problem, the California researchers leveraged existing image repositories, including over 2,000 histopathology images from The Cancer Genome Atlas (TCGA). With machine learning, they designed a tool that can scrutinize tumor cells for a wide array of cancer specific traits, including those that cannot be discerned by the human eye. Those features include straightforward elements such as cell size and shape, as well as more subtle characteristics, like the size and shape of cell nuclei, cell texture, and the relative positions of neighboring cells.
The researchers’ method accurately predicted survival (long-term versus short term) in patients diagnosed with stage 1 adenocarcinoma or squamous cell carcinoma. While lung and other tumors are currently classified according to grade (how irregular the tumor cells appear under the microscope) and stage (if and how far it has spread beyond its original anatomic location), this system is known to be imperfect. There can be wide variability within a single stage or grade of tumor. For example, in the TCGA cohort, over half of stage 1 adenocarcinoma patients died within 5 years of diagnosis, while about 15% lived for over a decade. This work provides a glimpse of a new generation of AI-based, data-driven methods in pathology, that can help clinicians tame complexity and make more precise diagnoses and outcome predictions for their patients.
For more information about Dr. Golden’s research, please contact Partners HealthCare Innovation by clicking here.
Most Recent Posts:
Exploring AI’s Impact on Industry, left to right: panelists, Marc Succi, MD, Daniela Rus and Paul English,…
Lilit Garibyan, MD, PhD, a dermatologist at the Wellman Center for Photomedicine and an associate professor of…
The Inflation Reduction Act: Responses From Experts on Implications for Biopharmaceutical Innovation
September 21, 2023—Two panel discussions on the Inflation Reduction Act at the World Medical Innovation Forum in…