Advanced computer-aided detection system exhibits no more false positives than experienced endoscopists in an...

The presence of computer-aided detection (CADe) false positives (FPs) may pose a hurdle to endoscopists' confidence in, and acceptance of this technology. However, endoscopists also have FPs, which refer to the fact that the object they initially suspected was ultimately confirmed not to be a polyp. The objective of this study was to compare the FPs and overall diagnostic performance between CADe and experienced endoscopists through an image-based study.

We prospectively enrolled 116 consecutive patients undergoing colonoscopy and diagnosed with pathologically confirmed small colon polyp, at Sichuan Provincial People's Hospital. Images of polyps and polyp-free areas in an environment of colon inflation and deflation were taken. Randomly arranged images were labeled by CADe system and a group of experienced endoscopists. Sensitivity, specificity, and accuracy were compared between the CADe system and endoscopists.

The study comprised 464 images from 116 polyps and poly-free areas. The CADe system and endoscopists exhibited sensitivity, specificity, and accuracy of 90.09% vs. 87.56% (P = 0.46), 86.64% vs. 85.41% (P = 0.79), and 88.36% vs. 86.48% (P = 0.43), respectively, across all labeled images. In the test set with images taken in fully inflated lumen, sensitivity, specificity, and accuracy for the CADe system and endoscopists were 92.24% vs. 92.61% (P = 1.00), 95.69% vs. 86.70% (P = 0.04), and 93.97% vs. 89.66% (P = 0.13), respectively. In the test set with images taken in deflated lumen, sensitivity, specificity, and accuracy for the CADe system and endoscopists were 87.93% vs. 82.51% (P = 0.35), 77.59% vs. 84.11% (P = 0.24), and 82.76% vs. 83.31% (P = 1.00).

The results indicate that a state-of-the-art CADe system exhibits a similar number of FPs as experienced endoscopists, while achieving high polyp sensitivity. Future study efforts to mitigate CADe false-positive distractions should focus not just on minimizing the false-positive rate, but also on educational, behavioral, or user-interface innovations to optimize CADe usability.