AI Model May Help Guide Cancer Diagnosis and Treatment

Researchers have developed an AI model that can perform an array of diagnostic tasks for multiple forms of cancer, improving upon existing AI systems. It has been likened to ChatGPT in terms of its flexibility across multiple cancer types.

Their study, published in Nature, shows how the new technology could predict prognosis and treatment response across 19 types of cancer.

Building upon AI systems for cancer

An important part of cancer diagnosis and treatment is histopathology, where highly trained experts analyze tumors under a microscope to look for features that can help doctors manage a patient’s care.

AI and machine learning techniques are being developed to help or even attempt to reproduce histopathologists' work. Such models can perform various functions like detecting cancer cells, predicting patient survival and identifying novel insights into tumor behavior.

But so far, research has focused on developing separate highly specialized models for each diagnostic task.

“Millions of cancer patients worldwide do not have access to the diagnostic expertise of expert pathologists,” Dr. Kun-Hsing Yu told Technology Networks. “In addition, standard pathology evaluations cannot reliably predict patient survival outcomes.”

Yu and colleagues developed their model to perform an array of different diagnostic tasks across 19 different cancer types.

“Although a few AI methods for analyzing pathology images have been proposed, we have observed significant performance declines when applying these methods to samples from different hospitals. This critical gap inspired my team to create a robust and generalizable AI system for pathology image analysis,” said Yu.

How to train your AI

The researchers named the new model CHIEF (Clinical Histopathology and Imaging Evaluation Foundation). It is the first to predict patient outcomes and validate them across international groups of patients.

“Simply put, the CHIEF model reviewed 44 terabytes of pathology imaging data and distilled the patterns most useful for representing these images,” Yu explained.

The AI works by reading digital images of tumors on microscope slides and predicts the tumor’s molecular profile by detecting the cancer cells’ features. These features can forecast a patient’s response to treatments and therapies such as surgery, chemotherapy, radiotherapy and immunotherapies.

“This approach allows CHIEF to extract biologically important signals from high-resolution pathology images, resulting in substantially higher performance and greater reliability than existing technologies," Yu explained.

Once CHIEF was trained on this data, the researchers next assessed its performance by testing it on over 19,000 whole-slide images collected across 24 hospitals worldwide.

They found CHIEF outperformed other state-of-the-art AI methods by up to 36% on tasks such as detecting cancer cells, identifying where tumors originated and predicting patient outcomes.

In cancer detection, CHIEF achieved nearly 94% accuracy across 15 datasets containing 11 cancer types – significantly better than existing approaches. It also had over 90% accuracy when examining previously unseen slides from surgically removed colon, lung, breast, endometrium and cervical tumors.

With whole-slide images, CHIEF was over 70% accurate in identifying mutations in 54 genes commonly altered in cancer. It also predicted the presence of mutations – in 18 genes spanning 15 anatomical sites – that are linked to responses to FDA-approved drugs.

CHIEF was also effective at predicting the survival of patients. In all cancer types and all patient groups studied, CHIEF distinguished patients with longer-term survival from those with shorter-term survival – outperforming other models by 8%.

This suggests that the AI can be used across a variety of clinical settings and has the potential to help smooth the path to precision medicine by identifying patients that won’t respond well to standard therapies.

Next steps before regulatory approval

Nonetheless, the technology is not without its limitations, as Yu explains: “One limitation of the study is that CHIEF primarily focused on common cancer types, and we did not have sufficient samples to evaluate its performance on rare cancers systematically. We are currently working on extending CHIEF’s capability to recognize rare cancers.”

There is also a sizable amount of work to be done to further validate the prediction model before it can be used in clinical settings. “We plan to prospectively validate CHIEF in clinical settings and pursue FDA approval for our AI-based prediction models,” Yu explained. “Once we obtain regulatory approval, clinicians across the country can use our AI tool to diagnose cancer and guide treatment plans.”

Reference: Wang X, Zhao J, Marostica E, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature. 2024:1-9. doi: 10.1038/s41586-024-07894-z

Dr. Kun-Hsing Yu was speaking to Dr. Sarah Whelan, Science Writer for Technology Networks.

About the interviewee:

Dr. Kun-Hsing Yu is an assistant professor in Harvard Medical School’s Department of Biomedical Informatics. His lab uses multiomics data from cancer patients to predict clinical phenotypes. He holds an MD from the National Taiwan University and a PhD in biomedical informatics from Stanford University.