Introduction: Probe-based confocal laser endomicroscopy (pCLE) is a promising modality for classifying polyp histology in vivo, but decision making in real-time is hampered by high-magnification targeting and by the learning curve for image interpretation. The aim of this study is to test the feasibility of a system combining the use of a low-magnification, wider field-of-view pCLE probe and a computer-assisted diagnosis (CAD) algorithm that automatically classifies colonic polyps.
Methods: This feasibility study utilized images of polyps from 26 patients who underwent colonoscopy with pCLE. The pCLE images were reviewed offline by two expert and five junior endoscopists blinded to index histopathology. A subset of images was used to train classification software based on the consensus of two GI histopathologists. Images were processed to extract image features as inputs to a linear support vector machine classifier. We compared the CAD algorithm's prediction accuracy against the classification accuracy of the endoscopists.
Results: We utilized 96 neoplastic and 93 non-neoplastic confocal images from 27 neoplastic and 20 non-neoplastic polyps. The CAD algorithm had sensitivity of 95%, specificity of 94%, and accuracy of 94%. The expert endoscopists had sensitivities of 98% and 95%, specificities of 98% and 96%, and accuracies of 98% and 96%, while the junior endoscopists had, on average, a sensitivity of 60%, specificity of 85%, and accuracy of 73%.
Conclusion: The CAD algorithm showed comparable performance to offline review by expert endoscopists and improved performance when compared to junior endoscopists and may be useful for assisting clinical decision making in real time.
Keywords: Colorectal cancer; Confocal laser endomicroscopy; Machine learning; Polyp histology.