Artificial Intelligence Model for Detection of Colorectal Cancer on Routine Abdominopelvic CT Examinations: A...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Colorectal cancers (CRCs) are sometimes missed by radiologists during routine abdominopelvic CT scans.
• There is a need for improved detection methods for CRC on standard CT examinations.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for detecting colorectal cancer (CRC) on abdominopelvic CT scans.
• To assess the AI model's performance in detecting CT-visible CRC without bowel preparation.

Main Methods:

• A retrospective study involving 3945 patients, with a training set and internal/external test sets.
• A transformer-based object detection network was adapted to create an AI model for automatic CRC detection.
• AI performance was evaluated using ROC analysis, sensitivity, and specificity, and compared to radiologist readers.

Main Results:

• The AI model achieved an AUC of 0.867 (internal) and 0.808 (external test sets).
• Sensitivity was 79.6% (internal) and 80.8% (external), with specificity of 91.2% (internal) and 90.9% (external).
• The AI model detected CRCs missed by radiologists and showed comparable or superior performance to human readers in some metrics.

Conclusions:

• The developed AI model demonstrates significant utility for the automated detection of colorectal cancer on routine abdominopelvic CT examinations.
• This AI tool has the potential to decrease the rate of missed CRCs in clinical practice.