Artificial intelligence and machine learning dramatically alter how medicine is practiced, and cancer detection is no exception, according to Toufic Kachaamy, MD of City of Hope Phoenix.
Since the first artificial intelligence (AI) enabled medical device received FDA approval in 1995 for cervical slide interpretation, there have been 521 FDA approvals provided for AI-powered devices as of May 2023.1 Many of these devices are for early cancer detection, an area of significant need since most cancers are diagnosed at a later stage. For most patients, an earlier diagnosis means a higher chance of positive outcomes such as cure, less need for systemic therapy and a higher chance of maintaining a good quality of life after cancer treatment.
While an extensive review of these is beyond the scope of one article, this article will summarize the major areas where AI and machine learning (ML) are currently being used and studied for early cancer detection.
The first area is large database analyses for identifying patients at risk for cancer or with early signs of cancer. These models analyze the electronic medical records, a structured digital database, and use pattern recognition and natural language processing to identify patients with specific characteristics. These include individuals with signs and symptoms suggestive of cancer; those at risk of cancer based on known risk factors; or specific health measures associated with cancer. For example, pancreatic cancer has a relatively low incidence but is still the fourth leading cause of cancer death. Because of the low incidence, screening the general population is neither practical nor cost-effective. ML can be used to analyze specific health outcomes such as new onset hyperglycemia2 and certain health data from questionnaires (3) to classify members of the population as high risk for pancreatic cancer. This allows the screened population to be "enriched with pancreatic cancer," thus making screening higher yield and more cost-effective at an earlier stage.
Another area leveraging AI and ML learning is image analyses. The human vision is best centrally, representing less than 3 degrees of the visual field. Peripheral vision has significantly less special resolution and is more suited for rapid movements and "big picture" analysis. In addition, "inattentional blindness" or missing significant findings when focused on a specific task is one of the vulnerabilities of humans, as demonstrated in the study that showed even experts missed a gorilla in a CT when searching for lung nodules.3 Machines are not susceptible to fatigue, distraction, blind spots or inattentional blindness. In a study that compared a deep learning algorithm to radiologist from the National Lung Screening trial, the algorithm performed better than the radiologist in detecting lung cancer on chest X-rays.4
AI algorithm analysis of histologic specimens can serve as an initial screening tool and an assistant as a real-time interactive interface during histological analysis.5 AI is capable of diagnosing cancer with high accuracy.6 It can accurately determine grades, such as the Gleason score for prostate cancer and identify lymph node metastasis.7 AI is also being explored in predicting gene mutations from histologic analysis. This has the potential of decreasing cost and improving time to analysis. Both are limitations in today's practice limiting universal gene analysis in cancer patients,8 but at the same time are gaining a role in precision cancer treatment.9
An excitingand up-and-coming area where AI and deep learning are the combination of the above such as combining large data analysis with pathology assessment and/ or image analyses. For example, using medical record analysis and CXR findings, deep learning was used to identify patients at high risk for lung cancer and who would benefit the most from lung cancer screening. This has great potential, especially since only 5% of patients eligible for lung cancer screening are currently being screened.10
Finally, the holy grail of cancer detection: blood-based multicancer detection tests, many of which are already available and in development, often use AI algorithms to develop, analyze and validate their test.11
It is hard to imagine an area of medicine that AI and ML will not impact. AI is unlikely, at least for the foreseeable future, to replace physicians. It will be used to enhance physician performance, improve accuracy and efficiency. However, it is essential to note that machine-human interaction is very complicated, and we are scratching the surface of this era. It is premature to assume that real-world outcomes will be like outcomes seen in trials. Any outcome that involves human analysis and final decision-making is affected by human performance. Training and studying human behavior are needed for human-machine interaction to produce optimal outcomes. For example, randomized controlled studies have shown increased polyp detection during colonoscopy using computer-aided detection or AI-based image analysis.12 However, real-life data did not show similar findings13 likely due to a difference in how AI impacts different endoscopists.
Artificial intelligence and machine learning dramatically alter how medicine is practiced, and cancer detection is no exception. Even in the medical world, where change is typically slower than in other disciplines, AI's pace of innovation is coming upon us quickly and, in certain instances, faster than many can grasp and adapt.