AI Model Detects Colorectal Cancer on Routine Noncontrast CT Scans

COCA reached Google News after detecting colorectal cancer with up to 88.2% sensitivity in consecutive patients, despite using routine noncontrast CT scans. The finding challenges a familiar limitation in radiology. These scans are common, yet clinicians have not considered them reliable colorectal cancer screening tools.

The model does not replace colonoscopy or create a new imaging appointment. Instead, it analyzes scans already collected for physical examinations, emergency care, outpatient visits, and hospital treatment. That makes the study less about another diagnostic algorithm and more about extracting overlooked clinical value from existing data.

The central contest is therefore clear. Traditional screening asks eligible people to complete a dedicated test. COCA represents opportunistic screening, which searches an existing examination for a disease unrelated to its original purpose.

Researchers reported strong retrospective results across multiple institutions and clinical settings. However, the model has not yet established that its alerts improve patient outcomes in prospective clinical practice. Regulatory review, workflow integration, follow-up capacity, and false-positive management remain unresolved.

Google News Attention Follows a Large Multicenter Study

The important change is not simply that AI recognized cancer, but that it found cancer on ordinary scans collected without colorectal screening preparation.

The study appeared online in Annals of Oncology on April 21, 2026, before publication in the journal’s August issue. Its authors included researchers from hospitals in China and the Czech Republic, alongside scientists affiliated with Alibaba’s DAMO Academy and Hupan Laboratory.

The researchers named the system COCA, short for Colorectal Cancer detection with AI. They developed it using scans from 1,321 patients with colorectal cancer and 1,357 normal controls at two centers.

These were noncontrast CT examinations, meaning clinicians did not inject an iodine-based contrast agent to make tissues and blood vessels easier to distinguish. The bowel was also not prepared as it would be for CT colonography.

The development design matters because colorectal tumors can be difficult to separate from nearby tissue on these images. Bowel contents, collapsed intestinal segments, and limited soft-tissue contrast can all obscure abnormalities.

COCA first localizes the colorectal region and then assesses it for cancer. The model combines lesion segmentation, which marks suspected tissue, with patient-level classification, which estimates whether the scan contains colorectal cancer.

Its mixed-supervised training strategy used different levels of annotation. One center supplied both patient labels and manually outlined tumor regions. Another supplied patient labels without the same complete voxel-level marking.

That approach allowed the system to learn from more scans without requiring radiologists to outline every tumor in three dimensions. Such outlining can consume substantial specialist time and often limits medical imaging datasets.

The investigators tested COCA beyond the two development centers. Their multicenter study included external abdominal, pelvic, and chest CT datasets from several institutions.

Across 2,053 patients in six-center validation, the reported area under the receiver operating characteristic curve ranged from 0.967 to 0.996. This measure summarizes how well a model separates positive and negative cases across different thresholds.

The researchers also conducted a reader study with 10 radiologists. Participants ranged from residents with three to four years of experience to specialists with 13 to 21 years.

Each radiologist reviewed 299 cases without assistance and then reviewed them with COCA after a washout period of at least four weeks. The model’s assistance increased reader sensitivity by 20.4% and specificity by 5.4%, according to the study.

Sensitivity measures how often a system identifies patients who have the disease. Specificity measures how often it correctly dismisses patients who do not have it.

The Google News coverage highlighted another operational detail. COCA reportedly processes a complete CT volume in approximately 30 seconds, making automated background analysis technically plausible within a radiology workflow.

These results establish a credible research signal. They do not establish a finished screening service, because retrospective performance is only the first part of clinical adoption.

Routine Scans Create an Opportunistic Screening Route

COCA shifts the screening question from whether patients will schedule another test to whether hospitals can responsibly reuse scans they already possess.

Current colorectal cancer screening depends on dedicated procedures or samples. Colonoscopy directly examines the bowel, while stool tests search for blood or molecular markers associated with cancer.

CT colonography also images the colon, but it requires bowel preparation and inflation of the colon. A routine noncontrast CT lacks those features and is generally ordered for another clinical question.

That difference creates COCA’s strategic value. A patient who enters an emergency department with abdominal pain might receive a CT scan for kidney stones, inflammation, or another suspected condition.

The radiologist must prioritize the clinical question that prompted the examination. An unprepared colon can receive less attention, particularly when the visible abnormality is small or subtle.

COCA can inspect that same scan in the background. It can then direct a radiologist toward a suspicious colorectal region rather than issue a final diagnosis independently.

This is opportunistic screening, the use of existing clinical data to search for an additional condition without arranging a separate screening encounter. Similar approaches already examine routine images for osteoporosis, cardiovascular risk, and body composition.

The route addresses a persistent participation problem. The 2026 screening guidelines recommend that average-risk adults begin screening at age 45.

The guidelines recommend continued screening through age 75 when life expectancy exceeds 10 years. Adults aged 76 through 85 should make individualized decisions with their health care providers.

Visual examinations and stool-based tests remain the preferred options. Positive results from non-colonoscopy tests should receive timely colonoscopy follow-up.

That last requirement also applies conceptually to COCA. A flagged scan would not confirm cancer by itself. It would create a reason for expert review and, when appropriate, diagnostic colonoscopy or another targeted examination.

The model therefore competes with screening nonparticipation, not with colonoscopy’s diagnostic authority. Its most useful patient is someone whose cancer appears incidentally on a scan but might otherwise remain unnoticed.

This distinction also prevents a common misunderstanding surrounding the Google News headline. Patients should not request medically unnecessary CT scans merely to obtain COCA analysis.

CT uses ionizing radiation, and the model has not been established as a reason to expose an otherwise unscreened person. Its proposed advantage comes from reusing scans that were already clinically justified.

The approach can also reach people outside standard screening patterns. Some patients receiving emergency or inpatient imaging have not completed recommended colorectal screening.

Others are younger than the traditional screening population or have barriers that delay planned testing. An automated review layer could flag a visible cancer without waiting for those barriers to disappear.

Hospitals would gain this opportunity without changing the scanner or acquiring a second image. Yet they would still need software integration, radiologist oversight, referral protocols, and capacity for follow-up procedures.

That operational burden leads to the main tension. An abundant source of imaging data can create more screening opportunities, but every algorithmic alert creates a clinical responsibility.

COCA’s Mechanism Targets the Parts Radiologists Miss

COCA combines anatomical localization with cancer classification, giving it a narrower and more interpretable task than a general image-recognition system.

A raw CT volume contains many organs, tissues, and possible abnormalities. Searching every voxel for an uncommon colorectal tumor would expose a model to considerable visual noise.

COCA handles that challenge in two stages. A U-Net segmentation network first identifies the colorectal region, reducing the area that the diagnostic component must evaluate.

U-Net is an encoder-decoder architecture that compresses an image into learned features and then reconstructs spatial detail. Skip connections preserve location information as the network processes the scan.

The second stage combines tumor segmentation with whole-patient classification. Multiscale image features are pooled into a representation used to estimate whether colorectal cancer is present.

The model also uses the suspected tumor volume when producing its final assessment. This design can show a radiologist where the system found concern, rather than returning only an unexplained risk score.

That location information matters in clinical review. A highlighted bowel segment lets the reader compare the alert with wall thickening, narrowing, surrounding tissue changes, or other visible findings.

The study reports that COCA performed particularly well in the sigmoid colon and rectum. Those sites account for a substantial share of colorectal cancers but can be difficult to evaluate on routine CT.

The developers also enriched training batches with smaller tumors. This step addressed the natural tendency of image classifiers to learn more easily from large, visually obvious lesions.

Internal results still show the remaining difficulty. Sensitivity for tumors smaller than three centimeters was 82.4%, based on 78 tumors in that subgroup.

For stage I disease, sensitivity was 76.7% among 30 cases. Stage II sensitivity reached 96% among 99 cases.

Those subgroup sizes are much smaller than the full validation population. Their confidence intervals are consequently wider, so the results should not be treated as exact estimates for every hospital.

The model’s colorectal cancer segmentation Dice score was 0.558. A Dice score measures spatial overlap between the model’s marked region and the reference annotation.

That result is useful but imperfect. The system can classify a scan correctly without tracing every tumor boundary precisely, which explains why detection performance can exceed segmentation accuracy.

The researchers also compared paired noncontrast and venous-phase contrast-enhanced scans at one site. Reported AUC values were 0.979 and 0.980, respectively.

That close result supports the proposed mechanism. COCA appears able to extract relevant patterns even when ordinary visual contrast is limited.

Adding 10 clinical risk factors, including age, sex, and family history, did not significantly improve the baseline model. The image signal drove the reported performance.

A separate routine CT model published earlier used a transformer-based detection network on contrast-enhanced examinations without bowel preparation. That system focused on CT-visible colorectal cancers in routine abdominopelvic imaging.

The comparison shows an emerging research route rather than a single isolated experiment. Researchers are testing whether software can turn general-purpose CT archives into secondary cancer-detection channels.

COCA extends that route to noncontrast scans and partial views from chest examinations. Its mechanism is therefore meaningful, but its clinical value depends on results from consecutive real patients.

Strong Real-World Numbers Still Leave Clinical Risks

The consecutive cohorts strengthen COCA’s case, but they also expose the difference between impressive accuracy and a dependable screening program.

The researchers evaluated two versions of the model in 27,433 consecutive patients across two real-world cohorts. These patients came from physical examinations, emergency departments, outpatient clinics, and inpatient services.

The first cohort included 9,014 patients. COCA recorded 88.2% sensitivity and 99.5% specificity.

The second external cohort included 18,419 patients. An iteratively improved version reached 86.6% sensitivity, 99.8% specificity, and a positive predictive value of 63.4%.

Positive predictive value answers a practical question. Among people flagged by the model, it measures how many actually had colorectal cancer.

A 63.4% value is substantial for opportunistic detection, but it also means some positive alerts did not represent cancer. Those alerts still demand review and potentially additional testing.

The researchers reported that some false positives involved nonmalignant abnormalities that could merit clinical attention. That can be useful, but it can also expand the model’s intended role.

A system designed to detect colorectal cancer needs a clear policy for incidental noncancer findings. Otherwise, hospitals risk uncertain responsibilities, inconsistent referrals, and avoidable patient anxiety.

False negatives matter just as much. Sensitivity of 86.6% means the model missed a portion of cancers in that cohort.

Patients and clinicians cannot safely interpret a negative COCA result as proof that cancer is absent. Routine CT also remains less suitable for detecting precancerous polyps than dedicated screening methods.

The model reportedly identified five cancers that clinicians had initially missed. One emergency patient returned for another scan one year later, but the tumor was not diagnosed until two years after the first examination.

That case illustrates the promise clearly. It also comes from retrospective review, where researchers already know which patients eventually received a diagnosis.

Prospective deployment changes the problem. Alerts arrive amid active workloads, incomplete records, competing emergencies, and limited follow-up capacity.

Radiologists must decide whether a marked region is genuinely suspicious. Referring clinicians must receive the result, explain it, and arrange confirmation.

Hospitals must also determine how quickly the algorithm runs and whether it delays image availability. They need plans for outages, software updates, performance monitoring, and cases with limited anatomy.

Population differences present another uncertainty. Development data came primarily from Chinese medical centers, with international testing that included a Czech institution.

That is a stronger design than single-center validation. It still does not guarantee identical performance across North American populations, scanner fleets, imaging protocols, or referral patterns.

Disease prevalence also changes positive predictive value. A model can maintain sensitivity and specificity while producing a different proportion of useful alerts in another clinical population.

The reader experiment has limitations too. Radiologists reviewed the same selected cases in two sessions, and they knew the task centered on colorectal cancer.

Daily practice is broader and less predictable. Readers search for many diseases while responding to each patient’s immediate clinical question.

The published report describes planned prospective work involving integration with picture archiving and communication systems, commonly called PACS.

That next phase is essential. It must show whether the model changes reporting, confirmation, treatment timing, and patient outcomes without producing unacceptable downstream harm.

Regulation and Workflow Will Decide Whether COCA Scales

COCA’s next test is not another retrospective accuracy benchmark, but controlled deployment inside real clinical pathways.

The researchers say the system can run through local Docker installation or secure cloud access after regulatory clearance. Both options raise practical questions for hospitals.

A local installation can keep imaging data within institutional infrastructure. It also requires hardware planning, software maintenance, cybersecurity controls, and reliable connections to imaging systems.

Cloud processing can simplify centralized updates. However, health systems must evaluate data transfer, privacy, latency, service availability, and contractual responsibility.

Neither route solves alert governance automatically. A hospital must decide who sees the alert, who owns the final interpretation, and how the result enters the medical record.

The system’s developers chose a high-specificity operating point for opportunistic screening. That decision reflects the cost of false alarms across a large volume of scans.

Even a small false-positive rate can generate significant work when software analyzes thousands of examinations. Radiologists must review each flagged case before patients enter a diagnostic pathway.

Changing the threshold creates a familiar tradeoff. Higher sensitivity finds more cancers but generally creates more false positives. Higher specificity reduces unnecessary alerts but allows more cancers to pass unnoticed.

A second configuration prioritized sensitivity above 95% during model development. The study explored that version for average-risk populations and advanced precancerous lesions.

Those two configurations should not be blended into one headline number. Each serves a different clinical objective and produces a different balance of missed disease and follow-up burden.

North American adoption would also require the applicable regulatory process. The FDA device list identifies AI-enabled products authorized for marketing in the United States.

The agency says listed devices have met applicable premarket requirements for their intended uses. Study design, safety, effectiveness, and technological characteristics factor into that review.

A publication alone does not provide such authorization. COCA’s developers would need to define its intended users, patient population, inputs, outputs, and role in clinical decisions.

Regulators and hospitals would also need to understand how the model changes over time. The paper describes hard-example mining and continual learning used to produce COCA Plus.

Continual learning means improving a model with additional data after its original training. In medical use, uncontrolled change can undermine the evidence supporting an approved version.

A deployment plan therefore needs version control and validation boundaries. Hospitals must know whether an update changes sensitivity, specificity, calibration, or performance for particular scanners.

Monitoring cannot stop after installation. Teams need to examine performance drift, alert acceptance, missed cancers, false-positive referrals, and disparities across demographic groups.

Clinical integration also requires restraint in patient communication. A COCA flag indicates suspicion on an existing image, not a confirmed diagnosis.

The established screening framework remains relevant. The screening recommendation includes stool tests, colonoscopy, flexible sigmoidoscopy, and CT colonography for eligible adults.

Routine noncontrast CT with COCA is not currently listed as an equivalent screening strategy. Presenting it that way would overstate the evidence and risk discouraging proven screening.

The more defensible near-term role is secondary review. COCA can direct attention to scans that already exist while standard screening and diagnostic pathways retain authority.

Three Signals Will Show Whether the Headline Holds Up

Prospective outcomes, regulatory status, and follow-up performance will determine whether COCA becomes clinical infrastructure or remains a promising research result.

The first signal is prospective PACS integration. Researchers plan to run the model automatically on eligible routine CT scans and send flagged cases to radiologists.

A useful trial should report more than sensitivity and specificity. It should measure how often radiologists accept alerts, how quickly patients receive confirmation, and whether diagnoses occur earlier.

The trial should also reveal operational costs. Alert volume, additional reading time, colonoscopy referrals, patient contact failures, and unnecessary procedures all shape real clinical value.

Strong prospective results would support the study’s central claim. Poor follow-through would weaken it, even if the underlying model retains high retrospective accuracy.

The second signal is a clearly defined regulatory submission or authorization. That process would force the developers to specify precisely how clinicians should use COCA.

An authorization for triage would differ from authorization for screening or standalone diagnosis. The intended use determines what evidence hospitals and patients should expect.

Regulatory materials could also disclose performance by subgroup, scanner, institution, and clinical setting. Those details would help North American providers assess whether the evidence matches their populations.

Until that happens, Google News visibility should not be confused with regulatory readiness. The public headline describes a study, not an available standard of care.

The third signal is the quality of downstream confirmation. Hospitals must show what happens after an alert reaches a clinician.

The most meaningful metric is not simply how many scans COCA flags. It is how many patients complete appropriate follow-up and receive a correct diagnosis without excessive harm.

Positive predictive value should remain stable when prevalence changes. False positives should not overwhelm radiology teams or colonoscopy services.

Researchers should also track interval cancers, which are cancers diagnosed after a negative result and before the next expected screening opportunity. That measure can reveal clinically important misses.

Equity deserves attention within the same evaluation. Opportunistic screening only reaches people who already receive CT imaging, and access to imaging is not evenly distributed.

The approach could help patients who miss conventional screening but undergo emergency or inpatient imaging. It could also favor populations with greater access to advanced hospital care.

COCA’s strongest idea is data reuse. A routine scan collected for one reason can contain a second, clinically important signal.

Its strongest evidence is the consecutive real-world testing across 27,433 patients. The reported 86.6% to 88.2% sensitivity suggests the signal is not limited to a carefully balanced laboratory dataset.

Its largest weakness is the remaining gap between detection and improved care. Every useful flag needs expert confirmation, patient communication, and an effective diagnostic pathway.

Readers following this story through Google News should watch those three signals in order: prospective workflow results, regulatory clarity, and completed follow-up outcomes.

If all three hold, routine noncontrast CT could become a valuable secondary safety net for colorectal cancer. Until then, patients should continue recommended screening and discuss suitable options with their clinicians.