Cancer — 13

Both micro and macro tests such as whole-body MRI can help people; the tests have detected cancer in some healthy people that led to definitive treatment. However, is there a net benefit or harm in your case and generally in the population? We don’t yet know, and we won’t until randomized trials assess clinical outcomes. At least for the micro tests, such randomized trials have been initiated, which will fill gaps in the evidence. Less helpful, Prenuvo launched an observational study of one hundred thousand people undergoing whole-body MRI over the next ten years.

While those gaps are being worked out, we do know that exploiting deep-learning AI interpretation of many medical scans and pathology slides delivers more accurate diagnoses. For medical images, the two most studied are mammography and colonoscopy. The largest randomized trial in medical AI to date was performed in Sweden, assessing radiologists’ interpretation of mammograms in eighty thousand women participants, with or without AI. The use of AI increased screen-detected cancers by 15 percent and reduced the workload by 44 percent. In Denmark, among about sixty thousand women screened with mammography before and after AI was implemented, there was improved cancer detection and fewer false positives, and the radiologist’s workload was reduced by a third. Other large prospective studies have shown similar results. This has also been replicated in the real world, in Hungary, with 13 percent more breast cancer diagnoses being made with the combination of AI and radiologists. In the United States, more than two dozen AI algorithms have been FDA cleared or approved for breast cancer detection, one of which is RadNet’s algorithm that compared the radiologist’s interpretation to the radiologist plus their AI (which includes a 3D reconstruction) and showed a 14 percent increased detection rate. However, none of these algorithms have billing codes or are reimbursable by insurance providers. When this was offered at RadNet’s 350 imaging centers distributed throughout the United States, women were asked to pay an additional $40 out of pocket for their AI mammogram interpretation. There’s another big issue: we don’t yet know whether this boost in detection will change outcomes for the people getting diagnosed. The big Swedish multiyear follow-up randomized trial will clarify what difference such detection makes for breast cancer prevention.

Over thirty-three randomized trials of colonoscopy compared gastroenterologists working alone or accompanied by real-time machine vision to pick up tumors of the colon or rectum, known as adenomatous polyps. The clear result: use of AI reduced missing polyps by more than 50 percent. In parallel with mammography, diagnosis of polyps is not the same as altering the natural history of colon cancer. The implications for breast and colon cancer imaging extend to many other organs, such as picking up lung nodules on chest X-ray or CT scans that may be cancerous, interpreting a prostate MRI or ultrasound for prostate cancer, diagnosing likely kidney cancer from an abdominal CT, and detecting pancreatic cancer from an abdominal or chest CT scan. For these cancer-imaging studies, the results are very encouraging for improved accuracy of cancer detection.