Table 2 Diagnostic performance of AI system for detecting and displaying early stage colorectal cancers and precursor lesions in still images.

From: Development of a real-time endoscopic image diagnosis support system using deep learning technology in colonoscopy

 

Sensitivity*, (n) (95% CIs)

Specificity†, (n) (95% CIs)

With lesions

Without lesions

All lesions (752 lesions)

97.3%

(732/752) (95.9–98.4)

90.9%

(638/702) (88.5–92.9)

99.0%

(4094/4135) (98.7–99.3)

Polypoid lesions (640 lesions‡)

98.1%

(628/640) (96.8–99.0)

90.4%

(535/592) (87.7–92.6)

—

Superficial lesions (112 lesions‡)

92.9%

(104/112) (86.4–96.9)

95.9%

(93/97) (89.8–98.9)

—

  1. *Sensitivity was defined as AI correctly detected lesion number/number of all lesions; †Specificity was defined as AI negative image number/true lesion negative image number (images without lesions); Correct answer was defined when AI detect and display loci of lesion by flag when the all three observers didn’t detect any lesions outside the flag or no flag, or when AI detect and display no loci when the image shows truly no lesion. ‡Since 13 images included both polypoid and superficial lesion, they were excluded from the subgroup specificity analysis (with lesions).
Back to article page