AI breakthrough raises hopes for better cancer diagnosis

Stay informed with free updates

Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.

A new artificial intelligence foundation model can accurately detect multiple cancer types, assess treatments and predict survival rates, in the latest advance in medical diagnosis driven by the technology.

The model — known as “Chief” — is a breakthrough because of the breadth of tumours it can analyse and its capacity to predict outcomes for patients, its Harvard Medical School inventors say.

Chief highlights how AI has powered improvements in diagnostic techniques based on images, in part because it is able to spot the significance of features even an experienced human eye may miss.

“Our ambition was to create a nimble, versatile ChatGPT-like AI platform that can perform a broad range of cancer evaluation tasks,” said Kun-Hsing Yu, assistant professor of biomedical informatics at Harvard Medical School’s Blavatnik Institute. “Our model turned out to be very useful across multiple tasks related to cancer detection, prognosis, and treatment response across multiple cancers.”

While recent breakthroughs in AI have led to fears about the technology’s abuse, optimists have argued it also has the power to produce long-term benefits to humanity in fields such as medicine and climate science.

Chief, described in a paper published in Nature today, works by reading digital slides of tumour tissues. It was trained on 15mn unlabelled sections of images and then 60,000 whole slide images of tissues, covering 19 different cancers.

The idea was to make sure Chief could relate detailed changes in one region of tissue to its wider context, the researchers said. They tested its performance on almost 20,000 whole-slide images from 24 hospitals and patient cohorts around the world.

Chief outperformed other AI diagnostic methods by up to 36 per cent in detecting cancer cells, predicting patient outcomes, and identifying tumour origins and the presence of genetic patterns related to treatment response, the paper said. Unlike some other current models, it had the versatility to maintain its performance regardless of the techniques used to obtain and digitise the tumour cells, they added.

Chief showed overall accuracy of almost 94 per cent for cancer detection, rising to 96 per cent for oesophagus, stomach, colon and prostate tumours. Its ability to link tumour cell patterns to specific genomic aberrations could help suggest best treatments without the need for costly and slow DNA sequencing, the scientists suggested.

The model offered further revealing information about tissues surrounding tumours, including the presence of a greater number of immune cells in long-term cancer survivors versus those who died sooner, the paper said.  

If Chief and similar approaches are validated by further research, they could be used to “identify early on cancer patients who may benefit from experimental treatments targeting certain molecular variations”, including in countries where that isn’t currently done, Yu said.

AI models are proving an increasingly useful ally to medical professionals in the field of imaging, because of their speed and pattern-spotting ability. While still imperfect, they can be helpful in triage, as a second opinion or to generate insights a doctor may have overlooked or not know about.  

Chief appeared to be an important new pan-cancer tool in a growing field of diagnostic foundational AI models, said Professor Eric Topol, founder and director of the Scripps Research Translational Institute in California.

In April, Harvard Medical School researchers at Boston’s Brigham and Women’s Hospital announced two models — known as Uni and Conch — to read, interpret and classify microscopy slides from patient tissues. They showed good results on diagnostic tasks ranging from disease detection to organ transplant assessment, as well as showing some ability to identify new and rare conditions.

The evolving new AI foundation diagnostic models promised to “provide exceptional insight from whole slide images, including improving accuracy of diagnosis and prognosis”, Topol said.