AI Reads the Liver on Routine CT Scans to Flag Colorectal Cancer Before Symptoms Appear

Colorectal cancer kills roughly 900,000 people worldwide each year, yet screening rates remain stubbornly low partly because colonoscopy is invasive, resource-intensive, and widely avoided. A new study published in Frontiers in Oncology proposes an unexpected detour: skip looking at the colon on CT entirely and look at the liver instead. Using an AI model trained on radiomic features extracted from liver segmentations, researchers achieved a test area under the receiver operating characteristic curve (AUROC) of 0.810, a level of discriminative power that dwarfs the clinical-only baseline and opens a credible route to opportunistic colorectal cancer screening.

Liver Radiomic Features Predict Colorectal Neoplasia with 0.810 AUROC

The core finding is striking: subtle quantitative features of liver morphology, invisible to the naked eye, carry a measurable signal about colorectal neoplasia occurring elsewhere in the body. The best-performing model, an XGBoost gradient-boosting classifier trained on the 20 most informative radiomic features, reached an AUROC of 0.810 (95% CI: 0.767 to 0.837) on the held-out test cohort. After threshold optimization, that translated to a sensitivity of 74.1% and a specificity of 72.3% for distinguishing patients with colorectal neoplasia from those without. A clinical-only model using conventional patient data achieved an AUROC of just 0.457, barely better than chance, which underscores how much diagnostic signal lives in the imaging data that standard risk scores never capture.

The concept driving the approach is the gut-liver axis: the bidirectional biochemical and microbial crosstalk between the colon and the liver, mediated largely through the portal circulation. When colorectal neoplasia develops, that axis is disrupted in ways that may leave detectable imprints on liver tissue composition. Radiomics, the extraction of quantitative metrics such as texture, shape, and intensity distributions from three-dimensional organ segmentations, provides a way to quantify those subtle structural shifts from imaging data that hospitals are already generating.

Opportunistic Screening Could Catch High-Risk Patients During Routine Abdominal CTs

The clinical appeal is the word "opportunistic." Millions of abdominal CT scans are performed every year for unrelated reasons: suspected appendicitis, kidney stones, follow-up of aortic aneurysms. Each of those scans contains a liver that an AI pipeline could analyze in seconds, flagging patients who may warrant colonoscopy without requiring any additional imaging or procedures. Updated colorectal cancer screening guidelines from the U.S. Preventive Services Task Force already call for screening starting at age 45 given rising incidence in younger adults, yet uptake of colonoscopy remains far from universal. A passive, scan-embedded signal that prompts a referral could meaningfully close that gap for patients who would otherwise fall through.

The neoplasia category in this study is deliberately broad: 808 patients had either adenomas (423) or colorectal cancer (385), while 1,189 served as controls. That breadth matters clinically because adenomas, the pre-malignant polyps that precede most colorectal cancers, are the ideal intervention point. Catching the signal at the adenoma stage is precisely when treatment is simplest and outcomes are best.

How the Team Built and Validated the Radiomic Pipeline

The study drew on retrospective data from 1,997 patients who had both a colonoscopy and an abdominal CT at the same institution, with the dataset split into a training cohort of 1,397 and a test cohort of 600. Three-dimensional liver segmentations were generated from each CT, and the Radiomics Processing ToolKit extracted quantitative features capturing everything from voxel-level intensity statistics to higher-order texture descriptors. Five machine learning architectures were trained using five-fold cross-validation, with XGBoost emerging as the top performer across the 20 most informative features selected through recursive feature selection.

This kind of AI-driven organ analysis sits at the frontier of what radiology is now capable of. Research on AI in whole-body cancer detection has shown that CT scans harbor information about tumors in unexpected locations, and AI integration across radiological cancer imaging has already reached or surpassed radiologist-level performance for lung and breast lesion detection in controlled settings. Applying a similar paradigm to an organ that serves as an indirect readout of colon pathology is a logical extension of that trajectory.

Single-Center Design and Subclassification Limits Temper the Results

The limitations here are real and the authors acknowledge them directly. This is a single-center, retrospective, proof-of-concept study. All 1,997 patients came from one institution, which means the radiomic features could be partially capturing scanner-specific or protocol-specific artifacts rather than true biological signal. Generalizability to other hospitals, scanner manufacturers, and patient populations remains unproven.

Subclassifying neoplasia into adenoma versus colorectal cancer was also harder: that model achieved an AUROC of only 0.674, considerably lower than the overall neoplasia-versus-control discrimination. This matters clinically because the management implications differ between a pre-cancerous polyp and an established cancer. Research on AI challenges in oncological imaging has consistently flagged the gap between controlled-study performance and real-world clinical specificity, and this study is no exception. A sensitivity and specificity in the low-to-mid 70s, while meaningful for a screening adjunct, would generate substantial false positives at population scale if deployed without further refinement.

The authors frame the work explicitly as a feasibility demonstration intended to motivate prospective, multicenter validation rather than immediate clinical deployment. That framing is appropriate given the design constraints.

Prospective Multicenter Validation Is the Critical Next Step

The researchers call for validation in larger, prospective, multicenter cohorts, and that requirement is non-negotiable before this approach could influence practice. Radiomic models are notoriously sensitive to acquisition parameters, and a feature that predicts neoplasia reliably on one scanner type may perform poorly on another without harmonization. Future work will also need to address whether the liver signal reflects systemic metabolic changes, microbiome-driven portal inflammation, or early subclinical metastatic seeding, any of which would carry different mechanistic and clinical implications.

If multicenter studies confirm the signal, the practical path forward would likely involve integrating the pipeline into existing CT reading workflows as a background process, similar to how AI tools now flag incidental pulmonary nodules. Radiologists would not need to change how they read scans; the system would simply append a colorectal risk flag to the report. That kind of lightweight integration, requiring no additional patient burden, is exactly what opportunistic screening is designed to be. The liver, it turns out, may have been quietly narrating the state of the colon all along.