Opportunistic AI detects colorectal cancer using routine, noncontrast CT
Researchers have developed a tool they say can detect colorectal cancer using noncontrast CT scans that have been completed for other clinical indications.
The current standard of care for diagnosing colorectal cancer is colonoscopy. However, due to the patient prep involved and the invasive nature of the exam, experts have been working to develop alternative methods. Among these, CT colonography stands out, as it is quick and noninvasive. Despite the benefits, adherence among eligible patients remains low. Researchers, however, havey developed a new AI assistant they believe can help address this issue.
The COlorectal Cancer detection with AI, or COCA, model is a cost-effective scalable solution that turns routine CT scans into opportunistic exams that can be used to proactively identify cancer. Experts believe COCA has large-scale potential in multiple imaging settings.
“Each year, tens of millions of routine abdominal and pelvic CT scans are performed for various clinical indications, such as trauma, nonspecific abdominal pain, or cancer staging, providing a unique opportunity for opportunistic [colorectal cancer] detection,” Z.Y. Liu, with the department of radiology at Guangdong Provincial People’s Hospital in China, and colleagues explained. “A largely overlooked aspect of these scans is their potential for CRC screening, as they often capture portions of the colorectum.”
In years past, noncontrast CT scans were not considered suitable for colorectal cancer detection; the emergence of AI has changed this. Now, AI-based tools have the ability to detect subtle imaging patterns on noncontrast exams that could be indicative of cancer. Their utility in colorectal cancer care, however, has been under explored.
To address this, the group developed and trained an AI-enabled program to detect signs of cancer on noncontrast abdomen/pelvis CT scans completed for a variety of clinical indications. This was achieved via a joint lesion segmentation and classification architecture, which was optimized with mixed-supervised learning. The AI tool was retrospectively tested by 10 radiologists tasked with identifying signs of colorectal cancer on scans from numerous settings—emergency, routine outpatient and inpatient—both with and without the assistance of AI.
On over 2,000 scans, COCA achieved an AUC ranging from 0.967 to 0.996 for cancer detection. It improved readers’ sensitivity by more than 20% and specificity by approximately 5%. When tested in a real-world cohort of more than 9,000 patients, COCA yielded a sensitivity of 88.2% and specificity of 99.5%. This performance was maintained in a second validation cohort of over 18,000 patients, suggesting that the program has the potential to fill in screening gaps observed in colorectal cancer screening.
“Compared with other noninvasive CRC screening methods, COCA offers a novel strategy that balances patient adherence and diagnostic performance, eliminating bowel preparation required by capsule colonoscopy and CT colonography,” the group noted.
The team signaled that they are continuing to validate the tool in clinical care and will fine-tune it as needed.
Read more about the findings here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.