Artificial Intelligence Model Detects Colorectal Cancer in Non-Contrast Computed Tomography

A newly developed deep learning model has successfully detected colorectal cancer in routinely performed non-contrast computed tomography scans with high real-world sensitivity and specificity.

Cancer Detection with Artificial Intelligence

Tens of millions of abdominal and pelvic computed tomography scans performed annually for various reasons such as trauma, abdominal pain, or staging are generally not examined with a focus on the colon. Trained to change this situation, the deep learning model showed high success in scans across six centers.

Screening Rates and Challenges

The US Preventive Services Task Force recommends colorectal cancer screening for all adults aged 45 to 75. However, screening compliance rates remain below target levels, and a significant portion of cancer deaths occurs in individuals who are not screened in a timely manner.

Non-contrast computed tomography scans were traditionally not considered suitable for the detection of colorectal tumors due to low soft tissue contrast and non-specific imaging characteristics.

Development of the COCA Model

The artificial intelligence-based colorectal cancer detection model named COCA was developed through a retrospective study. Optimized with weakly supervised learning, the model can process a computed tomography volume in approximately thirty seconds.

Multi-Center Validation and Success

In international validation processes involving numerous patients from six centers, the model significantly increased cancer detection sensitivity and specificity compared to radiologists.

Real-World Results and Future Plans

In real-world validation studies, the model achieved high sensitivity and specificity rates, identifying clinically missed cases in emergency and outpatient settings. The research team plans two prospective studies in the future.