At the Intersection of AI and Cancer Care
Every clinician knows that early detection is one of the most powerful tools we have in the war on cancer. Several computational tools are enabling us to make early detection a reality, at least in two specialties.
By John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform
If your clinician can be replaced by AI, your clinician should be replaced by AI – you must have heard this before. Diagnostic skill is a combination of empathy, active listening, respect, ethics, and knowledge. Since AI techniques can augment clinician knowledge, the key is finding the right balance between human effort and AI innovation.
Early detection is one of the keys to better prognosis, and that is especially true for cancers that typically remain asymptomatic until they reach late, life-threatening stages. Pancreatic ductal adenocarcinoma (PDAC), for instance, is almost always fatal because it is usually discovered too late. Median disease specific survival for Stage 1 PDAC is 26 months, compared to less than 5 months for stage IV. Ajit Goenka, M.D., and his colleagues at Mayo Clinic have developed several machine-learning based models that enabled them to recognize these tumors during a CT scan several months before the cancer became symptomatic. They performed volumetric segmentation on prediagnostic CT scans of pancreatic tissue and compared it segmentation data from CT scans of normal organs. They used several models in their retrospective analysis, including K-Nearest Neighbor, Support Vector Machine (SVM), Random Forest analysis, and XGBoost.
The median time from prediagnostic CT to cancer diagnosis was 386 days, with Support Vector Machine generating the most sensitive results and an area under the curve (AUC) reading of 0.98. AUC for diagnoses performed by human radiologists was only 0.66. The early detection would have given clinicians adequate time to cure the cancers surgically. The researchers hope to follow up the results of their case-control study with a more definitive prospective trial.
Using AI-based algorithms in the detection of colorectal cancer has proven even more promising. To date, there have been more than 10 randomized clinical trials that found AI-assisted polyp detection—often referred to as computer-assisted detection (CADe)—increases adenoma detection rates and decreases adenoma miss rates among endoscopists performing colonoscopes. These studies have been conducted in a variety of countries and diverse patient populations. Some trials have found AI-based algorithms can detect very small polyps (5 mm or smaller). In the latest Gastroenterology report, investigators found that CADe increased the detection of polyps between 5 and 10 mm.
The researchers evaluated over 1,300 patients who had either a screening or surveillance colonoscopy across 5 clinical sites in the United States. “The use of CADe provided a significant increase in APC [adenomas per colonoscopy] (1.05 compared with 0.83 in standard colonoscopy, which represented a relative increase of 27% with a P value of .002). Notably, this benefit in APC was not accompanied by an increased number of polypectomies of non-neoplastic lesions.” The CADe device used in the new Gastroenterology study--SKOUT; Iterative Scopes, Cambridge, MA*--relied on deep learning to create a visual marking box around the suspected lesion in the videos that the endoscopist can view on the monitor.
Another study by Mayo showed that AI reduced by twofold the rate at which precancerous polyps were missed in colorectal cancer screening.
As we have pointed out in previous blogs and journal papers, AI-based algorithms have several limitations; however, the above examples demonstrate that they are the future of medicine. The AI will never replace a skilled clinician, but it will empower them and become a seamless, behind-the-scenes part of health care.
*The mention of a software product does not imply endorsement by Mayo Clinic.
By John Halamka and Paul Cerrato — Mayo Clinic Platform_Connect is transforming how patient data is used to generate innovative diagnostic tools and treatment options.
By John Halamka and Paul Cerrato — What happens when ChatGPT-4 and a human cardiologist are asked to diagnose the same patient? The results are quite revealing.
By John Halamka and Paul Cerrato — Several thought leaders and stakeholders have joined forces to create GoodDx.org, a searchable database that has the potential to reduce the human suffering affecting millions of Americans.