Computer-aided diagnosis system for predicting deeply invasive colorectal cancer: a comparison with a prospective...

Computer-aided diagnosis (CAD) systems for predicting deeply invasive colorectal cancer (CRC) have been developed, but their clinical utility remains uncertain. This study aimed to compare the diagnostic accuracy of a CAD system with that of endoscopists using magnifying narrow-band imaging (M-NBI) and magnifying chromoendoscopy (MCE).

A total of 303 lesions obtained from a prospective multicenter study were analyzed. CAD diagnoses were classified as high- or low-confidence. The diagnostic accuracy of CAD was compared with that of M-NBI and MCE.

The diagnostic accuracy of the CAD system was 61.3%; sensitivity: 92.3%; specificity: 47.5%; positive predictive value (PPV): 95.4%; negative predictive value (NPV), and 89.1% overall accuracy. Among the 303 lesions, 193 (63.7%) were diagnosed with high confidence, with accuracy rates of 95.3% for high-confidence cases and 78.2% for low-confidence cases. For M-NBI, the sensitivity, specificity, PPV, NPV, and accuracy were 74.2%, 96.3%, 69.7%, 97.0%, and 94.1%, respectively. For MCE, the values were 83.9%, 95.2%, 66.7%, 98.1%, and 94.1%, respectively. The overall accuracy of the CAD system was lesser than that of M-NBI (p = 0.011) and MCE (p = 0.014). However, in high-confidence cases, the diagnostic performance of the CAD system was comparable to that of M-NBI and MCE.

In high-confidence cases, the diagnostic accuracy rates of the CAD system were comparable to those of M-NBI and MCE. However, in low-confidence cases, the accuracy of the CAD system was insufficient, and further evaluations using M-NBI and MCE are necessary.