Endocytoscopy and artificial intelligence-based diagnosis for early colorectal cancer.

Background and Aim:
Approximately 40% of colorectal cancers detected in a screening program are submucosal invasive (T1) colorectal cancer (oes-Zoutendijk E, et al. GUT 2018). Thus, we endoscopists often encountered T1 colorectal cancers; however precise depth prediction in community-based practice is still difficult. Namely, about 60% of T1 colorectal cancers were misdiagnosed as a benign adenoma (Vleugels J, et al. GUT 2020). To provide a precise diagnosis for all colonoscopists, we have been developing endocytoscopy and artificial intelligence (AI) for diagnosing early colorectal cancers.
Methods and results:
Endocytoscopy (CF-H290ECI, Olympus Corp.) is a contact type, endomicroscopy that enables in vivo cellular observation. Endocytoscopy with methylene-blue staining provides a similar image to pathological microscopic images by visualizing cell nuclei and glandular structure. Therefore, its diagnostic performance is close to the gold standard. We developed AI-assisted software (EndoBRAIN-Plus, Cybernet System Company) that is designed as providing three-class histopathological prediction (non-neoplastic, adenoma, and invasive cancer) by analyzing captured endocytoscopic still images. We conducted a retrospective analysis using 500 endocytoscopic images containing 20 colorectal lesions. The diagnostic sensitivity and specificity for invasive cancers were 91.8% and 97.3%, respectively.
Conclusion
AI may make it easier for even non-experts to diagnose early-stage colorectal cancer and enabling highly accurate and uniform diagnosis. We will introduce an AI-based diagnostic strategy for colorectal neoplasms in a coming conference.