Are you experiencing bottlenecks in your workflow when interpreting whole tissue slides? Machine learning (ML) can help address this slowdown, but the technology can be costly and slide annotations can be inconsistent, sometimes leading to less-than-accurate results.

With funding in part from NCI, researchers have a new method—Histomorphological Phenotype Learning (HPL)—that could help overcome these hurdles.

Unlike other ML techniques, which often rely on annotations, HPL automatically identifies “tiles” of similar-looking features and groups them into clusters. By comparing these clusters with an atlas of images, HPL is able to both classify cancer and predict how it will progress. HPL also produces an encyclopedia of tissue patterns across cancer types, allowing researchers to better understand the diverse morphologies in tumors.

According to Dr. John Le Quesne, a corresponding author on the paper, “Interpreting whole slide images has been a task of seemingly insurmountable complexity. With HPL, we’re able to assign an immediate quantitative measure to these images, letting us classify cancer and perform additional analysis. And it works in all cancer types.”

Dr. Aristotelis Tsirigos, a second corresponding author, credited The Cancer Genome Atlas (TCGA) with helping to show HPL’s prognostic abilities. He noted, “By adding images from 443 TCGA patients to our New York University cohort, we were able to show that HPL could predict when and if cancer would return, and, importantly, explain which histologic patterns increase the risk for each patient.”

Dr. Ke Yuan, also a corresponding author, added, “In the future, we plan to conduct training on larger and more diverse data sets and test our methodology in new ways. For example, we plan to apply HPL to data from clinical trials and small biopsy images to see if our model can identify patterns that help predict response to treatments and better diagnosis.”