Authors: Ishita Barua, M.D., Paulina Wieszczy, Ph.D., Shin-ei Kudo, M.D., Masashi Misawa, M.D., Øyvind Holme, M.D., Shraddha Gulati, M.D., Sophie Williams, M.D., Kensaku Mori, Ph.D., Hayato Itoh, Ph.D., Kazumi Takishima, M.D., Kenichi Mochizuki, M.D., Yuki Miyata, M.D., Kentaro Mochida, M.D., Yoshika Akimoto, M.D., Takanori Kuroki, M.D., Yuriko Morita, M.D., Osamu Shiina, M.D., Shun Kato, M.D., Tetsuo Nemoto, M.D., Bu Hayee, M.D., Mehul Patel, M.D., Nishmi Gunasingam, M.D., Alexandra Kent, M.D., Andrew Emmanuel, M.D., Carl Munck, M.D., Jens Aksel Nilsen, M.D., Stine Astrup Hvattum, M.D., Svein Oskar Frigstad, M.D., Petter Tandberg, M.D., Magnus Løberg, M.D., Mette Kalager, M.D., Amyn Haji, M.D., Michael Bretthauer, M.D., and Yuichi Mori, M.D.
BACKGROUND Artificial intelligence using computer-aided diagnosis (CADx) in real time with images acquired during colonoscopy may help colonoscopists distinguish between neoplastic polyps requiring removal and nonneoplastic polyps not requiring removal. In this study, we tested whether CADx analyzed images helped in this decision-making process.
METHODS We performed a multicenter clinical study comparing a novel CADx-system that uses real-time ultra-magnifying polyp visualization during colonoscopy with standard visual inspection of small (≤5 mm in diameter) polyps in the sigmoid colon and the rectum for optical diagnosis of neoplastic histology. After committing to a diagnosis (i.e., neoplastic, uncertain, or nonneoplastic), all imaged polyps were removed. The primary end point was sensitivity for neoplastic polyps by CADx and visual inspection, compared with histopathology. Secondary end points were specificity and colonoscopist confidence level in unaided optical diagnosis.
RESULTS We assessed 1289 individuals for eligibility at colonoscopy centers in Norway, the United Kingdom, and Japan. We detected 892 eligible polyps in 518 patients and included them in analyses: 359 were neoplastic and 533 were nonneoplastic. Sensitivity for the diagnosis of neoplastic polyps with standard visual inspection was 88.4% (95% confidence interval [CI], 84.3 to 91.5) compared with 90.4% (95% CI, 86.8 to 93.1) with CADx (P=0.33). Specificity was 83.1% (95% CI, 79.2 to 86.4) with standard visual inspection and 85.9% (95% CI, 82.3 to 88.8) with CADx. The proportion of polyp assessment with high confidence was 74.2% (95% CI, 70.9 to 77.3) with standard visual inspection versus 92.6% (95% CI, 90.6 to 94.3) with CADx.
CONCLUSIONS Real-time polyp assessment with CADx did not significantly increase the diagnostic sensitivity of neoplastic polyps during a colonoscopy compared with optical evaluation without CADx. (Funded by the Research Council of Norway [Norges Forskningsråd], the Norwegian Cancer Society [Kreftforeningen], and the Japan Society for the Promotion of Science; UMIN number, UMIN000035213.)