AI Detects Multiple Cancers With High Accuracy

Versatile AI digital pathology tool outperforms state-of-the-art deep learning.

Reviewed by Davia Sills

A peer-reviewed Harvard Medical School study recently published in Nature shows how a new artificial intelligence (AI) digital pathology tool called CHIEF (Clinical Histopathology Imaging Evaluation Foundation) outperforms state-of-the-art deep learning methods and detects multiple cancer types with up to 96 percent accuracy.

Cancer is one of the leading causes of death, responsible for an estimated 9.3 million deaths worldwide each year, according to the World Health Organization (WHO).

Early detection of cancer may result in better outcomes. The five-year survival rate for many cancers is more than 90 percent when detected early, according to the American Cancer Society. According to the National Cancer Institute’s survival statistics in the SEER (Surveillance, Epidemiology, and End Results) database, the five-year relative survival rates for localized stage cancers are nearly 100 percent for prostate, 99 percent for breast, 91 percent for colorectal, 91 percent for cervical, and 65 percent for lung cancers.

According to Grand View Research, AI in oncology is expected to reach USD 19 billion in revenue globally by 2030, with a compound annual growth rate of 28.9 percent from 2024 to 2030. The predictive capabilities of artificial intelligence machine learning offer hope as a potential diagnostic tool to assist human clinicians seeking to detect cancer early.

“CHIEF provides a generalizable foundation for efficient digital pathology evaluation for patients with cancer,” wrote corresponding author Kun-Hsing Yu, M.D., Ph.D., a recipient of the 2024 Harvard Medical School Dean’s Innovation Award and assistant professor of biomedical informatics at Harvard Medical School, along with his research colleagues.

What sets CHIEF apart from standard AI methods for histopathology image analyses is that its architecture was designed to be more generalizable rather than optimized for a single diagnostic task.

For this AI study, the researchers sourced their data from 16 pathology datasets and pretrained their solution in two steps. First, CHIEF was trained on 15 million unlabeled images grouped into areas of interest. Next, the scientists used data from 60,530 whole-slide pathology images across 19 anatomical sites (brain, pancreas, lung, breast, prostate, testicular, skin, soft tissue, adrenal gland colorectal, bladder, stomach, esophageal, kidney, thyroid, cervical, uterine, ovarian, liver) from 14 study cohorts to pretrain the AI model.

“Through pretraining on 44 terabytes of high-resolution pathology imaging datasets, CHIEF extracted microscopic representations useful for cancer cell detection, tumor origin identification, molecular profile characterization, and prognostic prediction,” the researchers wrote.

The AI model was validated with more than 19,400 whole-slide image data from 32 independent slide sets from 24 hospitals and international cohorts.

The researchers reported that their AI model had achieved a high accuracy rate of 96 percent in cancer detection for multiple cancer types, such as prostate, colon, stomach, and esophageal, when evaluated on five biopsy datasets from independent cohorts. When CHIEF was tested with new images that the model had not seen before on surgically removed tumors from the breast, lung, colon, cervix, and endometrium, it achieved over 90 percent accuracy.

In the task of predicting gene mutations, CHIEF achieved 96 percent for finding an EZH2 gene mutation common in diffuse, large B-cell lymphoma blood cancer, 91 percent accuracy for the NTRK1 gene mutation found in head and neck cancers, and 89 percent accuracy for the BRAF gene mutation in thyroid cancer.

For the task of predicting patient survival, CHIEF was able to predict which patients would survive longer for all the cancer types and patient groups studied and outperformed existing models by 10 percent in predicting later-stage cancers.

The scientists also enabled CHIEF to identify and highlight areas of interest on images, like a heatmap to draw attention to critical areas for survival prediction.

“Overall, CHIEF outperformed the state-of-the-art deep learning methods by up to 36.1 percent, showing its ability to address domain shifts observed in samples from diverse populations and processed by different slide preparation methods,” the researchers reported.

At the intersect of artificial intelligence and pathology are trailblazing research studies such as this one that are rapidly advancing precision oncology towards a better future.

Copyright © 2024 Cami Rosso. All rights reserved.