New AI model shows promise for cancer diagnosis

Abstract

A new ChatGPT-like artificial intelligence (AI) model developed by researchers at Harvard Medical School outperforms other state-of-the-art AI methods by up to 36% in an array of diagnostic tasks across multiple forms of cancer. These tasks include the detection of cancer cells, the identification of a tumor’s origin, the prediction of patient outcomes, and the identification of the presence of genes and DNA patterns associated with treatment response.^{1, 2}

The Clinical Histopathology Imaging Evaluation Foundation (CHIEF) model achieved nearly 94% accuracy in cancer detection across 15 different databases containing 11 cancer types. Its accuracy increased to 96% when it was based on five biopsy data sets for multiple cancer types, which included esophageal, stomach, colon, and prostate cancers, and achieved more than 90% accuracy when it was based on previously unseen slides from surgically removed tumors of multiple cancers, which included colon, lung, breast, endometrial, and cervical cancers.

The model was trained on more than 60,000 whole-slide pathology images spanning 19 anatomical sites, and it works by reading digital slides of tumor tissues to detect cancer cells and predict a tumor’s molecular profile.

For predicting tumors’ molecular profiles, CHIEF successfully identified several important genes associated with cancer growth and suppression and predicted key genetic mutations related to a tumor’s potential response to targeted therapy. When tested on US Food and Drug Administration–approved targeted therapies (across 18 genes in 15 anatomic sites), the model was 96% accurate in detecting EZH2 in diffuse large B-cell lymphoma, 89% accurate in detecting BRAF in thyroid cancer, and 91% accurate in detecting NTRK1 in head and neck cancers.

Using tumor histopathology images obtained at initial diagnosis, CHIEF also was able to predict patient survival and outperformed other AI models by 8% in its ability to distinguish between patients with longer- and shorter-term survival (for all cancer types) and by 10% in its ability to distinguish survival rates in patients with advanced cancers.

“The performance that the model was able to demonstrate across a diverse set of types and tasks was very impressive,” says Fei Wang, MD, professor of population health sciences and founding director of the Institute of Artificial Intelligence for Digital Health at Weill Cornell Medicine, who thinks that AI has huge potential to augment clinical tasks and improve patient care.

However, he underscores the need for rigorous prospective evaluations through clinical trials to better understand how these types of models work in the real world.

“The current model has not been tested in real-world clinical care settings, so it is not clear how robust the model will be with respect to real-world challenges, such as imaging quality, computing resource, communication speed, and practice variations,” he says.

As the model works only on pathology slides, Dr Wang says that it is important to see how it could be combined with other information modalities, such as genomics and clinical records as well as environmental exposures and behavioral changes.

Plans for refining the model are underway and include exposing the model to more molecular data, training the model to predict the benefits and adverse effects of novel and standard cancer treatments (including samples from premalignant tissues cells), and training the model on tissue samples from diseases other than cancer.