P0769 (S0326). - The Strengths of Artificial Intelligence and Water Exchange Complement the Weaknesses of Each Other to Yield Significantly Higher Additional Polyp Detection and Lower False Alarm Rates
Buddhist Dalin Tzu Chi General Hospital Chiayi, Chiayi, Taiwan (Republic of China)
Chia-Pei Tang, MD, MBA1, Yu-Hsi Hsieh, MD1, Tu-Liang Lin, PhD2, Felix W. Leung, MD3; 1Buddhist Dalin Tzu Chi General Hospital, Chiayi, Chiayi, Taiwan; 2National Chiayi University, Chiayi, Chiayi, Taiwan; 3VAGLAHS, David Geffen School of Medicine at UCLA, North Hills, CA
Introduction: The strengths and weaknesses of water exchange (WE) and artificial intelligence (AI) guided computer-aided detection (CADe) are summarized in Table 1. We test the hypothesis that the strengths of WE and AI complement the weaknesses of each other such that combining WE and AI gives rise to significantly higher additional polyp detection rate (APDR) and lower false alarm rate (FAR). Methods:
Methods: To detect colon polyps, we have developed a CADe system using convolutional neural network with transfer learning approach by dividing static polyp images (n = 5463) randomly into training (n = 4638) and testing (n = 825) datasets. We used this CADe model to evaluate edited videos of high-definition colonoscopy in a RCT that compared right colon ADR between air insufflation and WE. Only the withdrawal phase of the right colon was analyzed by the blinded reviewers. Two senior endoscopists analyzed the CADe-overlaid videos at the same time and reached consensus on whether additional polyps were detected. A detected artifact which was continuously traced by the system but not considered a polyp by the reviewers was defined as FA. The primary outcome was APDR. The secondary outcome was the FAR. Results:
Results: 123 (WE) and 122 (air insufflation) edited videos were analyzed. The CADe evaluation of video recordings was tabulated as APDR and FAR (Table 2). APDR was 19/123 (15.4%) and 5/122 (4.1%) for the WE and air insufflation group, respectively (P = 0.004). Among the additionally detected polyps (Figure 1A), 6 of them were characterized as adenoma by computer-aided diagnosis (CAD) (Figure 1B). The mean total number of polyp detection was increased in WE compared with air insufflation group (0.98 [0.11] vs. 0.61 [0.10], P = 0.006). The FAR (Figure 1C) for the WE group was significantly lower than that of the air insufflation group (5/123 [4.1%] vs.19/122 [15.6%], P = 0.002). The Boston Bowel Preparation Scale score during the original colonoscopy was significantly higher in the WE arm. Discussion:
Conclusions: The current results showed that CADe system revealed additional polyps, adenomas and false alarms in both study arms. A significantly higher APDR and a significantly lower FAR occurred with WE. AI promoted recognition of polyps that were missed due to inattention or inadequate inspection technique. The data support the hypothesis that the strengths of the two methods complement the weaknesses of each other in increasing polyp detection and decreasing false alarms.
Disclosures: Chia-Pei Tang indicated no relevant financial relationships. Yu-Hsi Hsieh indicated no relevant financial relationships. Tu-Liang Lin indicated no relevant financial relationships. Felix Leung indicated no relevant financial relationships.