The Jackson Laboratory – How one AI approach is helping pathologists diagnose and treat cancer patients

Back

The Jackson Laboratory (JAX) is using AI approaches to improve cancer classification and prediction of response to treatment based on histology images, a standard and abundant data type for cancer patients.

Pathologists have traditionally used such images to diagnose cancer patients and suggest treatment approaches. However, manual pathology review is time consuming, can vary between pathologists, and is often unable to effectively distinguish which treatment will be most successful. Over the last several years, JAX has adopted deep learning neural network approaches to identify the common features within such images to automate and extend what human pathologists can accomplish. Current approaches have not sufficiently explored how to distill an image into its most essential elements (how to mathematically “represent” an image). Common current approaches represent an image, for example, based on the counts of different types of cells, calculate the “average” behavior within an image, or try to identify particular regions where attention should be directed. Such representation approaches have weaknesses that limit the ability to make clinically useful predictions from images. 

The Chuang Lab has developed an approach that mimics the way pathologists evaluate images, i.e. by taking into account the full distribution of behaviors within an image. The team not only considers the average behavior within an image, they also review the relative proportions of an image exhibiting different behaviors. This corresponds to the intuitive, interpretable way that human pathologists assess the relative prevalence of cancer tissue, dead tissue, and inflammatory tissue within an image. This approach is as accurate as much more complex neural network approaches and has major advantages in speed (often 1000x faster) and interpretability. These are critical issues for putting digital pathology into practice, as interpretability and speed are needed to assure clinicians that AI predictions are meaningful and actionable.

Learn more about the work of the Chuang Lab on jax.org.

 

Submitted by Jeff Chuang, Ph.D., professor, the Jackson Laboratory (JAX)