Each week we find a new topic for our readers to learn about in our AI Education column.
The term “cancer moonshot” was coined by President Joe Biden to encompass a set of public policy proposals aimed at a very personal target: treating and curing cancer, the disease that claimed the life of Beau Biden, one of his sons. I feel like a real, tangible battle against cancer is something that, done right, could gather bipartisan support, because practically everyone has someone they love who has died or battled the disease.
If there is going to be a “moonshot” for cancer, then AI is going to be like the Saturn V rocket that propelled the Apollo astronauts on their own non-metaphorical moonshot.
How We Got Here
The Harvard Medical School’s Ekaterina Pesheva published a story on Wednesday about CHIEF, a “versatile, ChatGPT-like AI model capable of performing an array of diagnostic tasks across multiple forms of cancers.”
Despite huge advances in not just technology, but our understanding of the disease itself, diagnosis for many cancers is reactive and old-fashioned. A patient comes to a doctor complaining of changes—symptoms like new and persistent discomfort or pain, fatigue, sweating without exertion. Based on the complaints, doctors order blood tests and images. The results of those tests and images are read and interpreted by doctors—which can still be a painstakingly manual process—and then, typically, some type of biopsy is performed to diagnose the cancer, including the specific type of cancer
One of the problems the medical system has with cancer is that in the manual processes, a lot can go wrong, and a lot can be missed. Some symptoms can be misinterpreted. Small tumors might go undetected. When it comes to treating and surviving cancer, early detection is key, and there are many opportunities for delayed diagnoses.
What CHIEF Does
Harvard’s new tool, Clinical Histopathology Imaging Evaluation Foundation, or CHEIF, analyzes tissue samples to accurately type cancer. CHIEF does this so well that it can help doctors create a treatment regimen specialized for the type of tumor and cancer, and personalized to the patient.
Not only that, but according to Pesheva’s story, “tool appears capable of generating novel insights — it identified specific tumor characteristics previously not known to be linked to patient survival.”
Just as large language models are trained on text, CHIEF was trained on medical imagery and pathology tissue slides—millions of them. Harvard’s tests found that it achieves a 94% accuracy in cancer detection, and can successfully predict a patient’s survival based on tumor images obtained at the time of diagnosis. Researchers are now training CHIEF to differentiate between healthy, pre-malignant and malignant tissues, as well as to identify non-cancerous diseases.
How Did We Connect AI To Diagnostics?
The healthcare system was slow to digitize. Silicon Valley entrepreneurs began moving into health care 30 years ago, but progress on digitizing diagnostic imagery, patient demographics and charting really didn’t break through until about 15 years ago.
Over the past 15 years, however, healthcare institutions have collected hundreds of millions of images and digitized tissue samples, lab results and lines of patient demographic data, all accessible within their computer systems.
The technology to pore through the tremendous volume of data they were collecting on communities and their health is even more recent. With the advent of AI, not only can researchers more easily use that data to find meaningful patterns and understand health, but they can also train algorithms to do a lot of the diagnostic dirty work at the front lines of healthcare.
CHIEF Is Not Working Alone
CHIEF is a promising diagnostic tool, but it is not the only time an AI model has been built to better detect cancer. Earlier this year, the National Institutes of Health, in conjunction with the Memorial Sloan Kettering Cancer Center, announced LORIS, an AI scoring system to help predict a patient’s response to immunotherapy.
Johns Hopkins researchers have developed an AI technology that can scan blood samples for DNA fragments to identify people more likely to have early-stage lung cancer.
The Mayo Clinic claims it has created a “new class” of AI it calls “hypothesis-driven” AI as a research tool to help clinicians and epidemiologists understand the intersecting causes of cancer, which could ultimately help in prevention, treatment and cure.
In the future, AI is going to help doctors detect cancer earlier, identify cancer faster, and as a result, successfully treat and cure more patients.
Where Else Does AI Come Into The Cancer Battle
CHIEF needs a tissue sample to do its work. Other AI models are being trained to look through imagery—CT scans, mammograms, x-rays, MRIs—to detect signs of cancer that doctors might have missed. On the other side of the equation, AI’s ability to read through piles of data to find a proverbial needle in a haystack promises to help rest the weary eyes of radiologists around the world.
Because models like CHIEF can accurately predict how different types of cancer will react to different traditional treatments, they can also be used to craft advanced treatments like immunotherapy, where the intervention has to be precise and specific to the patient and disease.
AI can also help doctors and patients optimize treatments better to reach their desired outcomes. So when a treatment team faces the dilemma of continuing medical treatment like chemotherapy or proceeding to surgery, predictive AI can help the patient and their doctors understand the consequences of each choice, allowing them to more quickly reach a decision.
