Accuracy of Computer-aided (CADx) System in Real-time Characterization of Colorectal Ulcerative Diseases

Colorectal Ulcers, Computer-aided System

The goal of this observational study is to test the diagnostic accuracy of the newly developed CADx system in predicting the histopathology of colorectal ulcers when compared to expert endoscopists. The main question it aims to answer are to demonstrate whether the newly developed CADx system has a high-level diagnostic accuracy in predicting characterization of colorectal ulcerative diseases. It is a multi-center study with two phases. The first retrospective phase is the development and validation of a CADx system by feature extraction from endoscopic photos and videos. The second prospective phase is the evaluation and comparison of the diagnostic accuracy between the CADx system, expert endoscopists and junior endoscopists.

In the first retrospective phase of study, our primary aim is to develop and validate a CADx system. Endoscopic photos and videos will be retrieved from existing database in the study centers (Nanfang Hospital). For each colorectal ulcer, different endoscopic views will be captures.Relevant baseline demographics, laboratory reports, imaging reports, endoscopy reports and histopathology results will be collected for analysis. The location, size and morphology of each colonic lesion will be recorded. The diagnosis of all colorectal ulcerative disease was comprehensively evaluated by independent pathologists and gastroenterologists. In our study, we will focus on the following subtypes of colorectal ulcerative lesions: 1)colorectal cancer (CA);2)Crohn's disease (CD);3)Ulcerative colitis (UC);4)intestinal tuberculosis (ITB);5) ischemic colitis (IC). All data will be de-identified before central processing to ensure confidentiality. A project-specific serial number will be used to represent each individual subject. All clinical data and de-identified endoscopic images will be kept confidential and will not be shared with any third party. A training cohort will be developed from majority of the included cases, followed by a validation cohort with the remaining cases. The endoscopic images and videos will be prepared to train the convoluted neural network and recurrent neural network by selecting appropriate regions of interest (ROI). Multiple ROI within the same colorectal ulcerative disease will be collected to reduce selection bias. Annotation and validation of endoscopic images will be performed by research team. The images will be further segmented into tiles of the same size for further processing. Deep learning algorithms will be applied to learn and extract features on the image and video data. We will develop the recurrent convolutional network to leverage the complementary information of visual and temporal features extracted from the video. Validation data are also created under the same principle which enable cross-validation for model accuracy. In the second prospective phase of study, we aim to compare the diagnostic accuracy between the CADx system, expert endoscopists and junior endoscopists. A set of test images and videos will be collected prospectively from other subjects, according to the previous eligibility criteria, followed by random allocation of computer-generated sequence. Two expert endoscopists (with more than 5 years of experience in colonoscopy and a total number of procedures more than 1,000) and two junior endoscopists (with less than 3 years of experience in colonoscopy and a total number of procedures less than 500), who are blinded to the final diagnostic result, will be invited to classify the test set images and videos according to the pre-defined subtypes. All endoscopists will assess the test set data independently in a real-time basis. On the other hand, the CADx system will scan the test set images and videos independently. The prediction of ulcer subtypes will be recorded. The formal diagnostic report after evaluation by independent pathologists and gastroenterologists will be regarded as the ground truth.

No related papers found yet.