AI Model Flags Colorectal Cancer on Routine Noncontrast CT Scans

Tens of millions of abdominal and pelvic CT scans are performed every year for trauma, vague abdominal pain, or staging, and almost none of them are read with the colon in mind. A deep learning model trained to change that detected colorectal cancer on routine noncontrast CT with an area under the curve of 0.967 to 0.996 across 6 centers.¹

Screening Adherence Gaps Left Room for an Opportunistic Approach

The US Preventive Services Task Force recommends colorectal cancer screening for all adults aged 45 to 75 years, yet uptake has lagged the target. In 2016, 26% of eligible US adults had never been screened for colorectal cancer, and in 2018, 31% were not up to date with screening.²

Existing options each carry a barrier. Colonoscopy is invasive, CT colonography requires bowel insufflation, capsule colonoscopy requires rigorous bowel preparation, stool DNA testing runs into reluctance to collect samples, and cell-free DNA blood tests have suboptimal sensitivity. Screening adherence remains below 60% against a target of 80%, and 76% of colorectal cancer deaths occur in people who were never screened in time.¹ Noncontrast CT has traditionally been considered unsuitable for detecting colorectal tumors because of low soft tissue contrast and nonspecific imaging features.

Multicenter Development and International Validation

The retrospective study developed colorectal cancer detection with AI (COCA), using 1321 patients with colorectal cancer and 1357 controls from 2 centers. The model used a joint lesion segmentation and classification architecture optimized with mixed-supervised learning, and it processes a CT volume in approximately 30 seconds.

Validation drew abdominal and pelvic CT data from 4 external centers and chest CT data from 4 centers. A reader study enrolled 10 radiologists of varying experience who interpreted noncontrast CT first without and then with model assistance. Real-world performance was assessed in 2 multiscenario cohorts totaling 27,433 consecutive patients presenting through physical examinations, emergency departments, outpatient clinics, and inpatient services.

Sensitivity Held Up in Consecutive Real-World Patients

In multicenter and international validation involving 2053 patients across 6 centers, the model achieved an area under the curve of 0.967 to 0.996. Compared with radiologists, it improved colorectal cancer detection sensitivity by 20.4% and specificity by 5.4%. Gains are concentrated in the sigmoid colon and rectum, which together account for roughly 40% of cases and are among the hardest regions to assess.

In the first real-world validation of 9014 consecutive patients, sensitivity was 88.2% and specificity 99.5%. In a second external real-world cohort of 18,419 consecutive patients, sensitivity was 86.6%, specificity 99.8%, and positive predictive value 63.4%. The model identified 5 clinically missed colorectal cancers across emergency, inpatient, and outpatient settings.

“COCA demonstrated robust performance across various clinical scenarios, including physical exams, emergency departments, outpatient, and inpatient settings, effectively preventing missed CRC diagnoses,” the authors wrote.

False Positives and Scope Limit Near-Term Use

The authors identified 1 principal limitation: the model detects colorectal cancer only and does not assess non-cancer pathology such as inflammatory bowel disease or benign lesions, so it cannot yet distinguish malignancy from the other abnormalities it flags. Many false positives corresponded to clinically relevant findings, including ulcerative colitis, Crohn disease, diverticulitis, and appendicitis, which warrant follow-up on their own terms.

At 99.5% specificity, roughly 50 false-positive results would arise per 10,000 individuals screened, and the authors proposed expert review of flagged cases before recall to contain that burden. Two prospective studies are planned, including a population-based screening program targeting asymptomatic adults aged 45 to 75 years. Regulatory clearance, patient consent for incidental findings, and secure integration into hospital picture archiving systems remain prerequisites, and cross-study comparisons with stool DNA and blood-based tests were nonrandomized and drawn from distinct populations.

References

1. Chen X, Qiu MY, Zhang JP, et al. Colorectal cancer detection using noncontrast CT and deep learning: a multicenter and international cohort study. Ann Oncol. 2026;37(8):1144-1156. doi:10.1016/j.annonc.2026.04.009
2. Davidson KW, Barry MJ, Mangione CM, et al; US Preventive Services Task Force. Screening for colorectal cancer: US Preventive Services Task Force recommendation statement. JAMA. 2021;325(19):1965-1977. doi:10.1001/jama.2021.6238