Deep Learning-Based Confocal Laser Microendoscopy Feature Atlas Construction and Its Application in Intelligent Diagnosis of Irritable Bowel Syndrome

Irritable Bowel Syndrome (IBS)

In this study, the healthy control group and IBS patients are taken as the research subjects, and CLE is applied to image and analyze the enrolled persons respectively, to derive and compare the characteristic microstructural images of the healthy control group with those of the IBS patients, to establish a diagnostic model of IBS by using deep learning and to evaluate the diagnostic efficacy of the model for IBS.Participants will undergo colonoscopy and confocal laser microendoscopy, and IBS patients will undergo targeted biopsy in CLE to observe suspicious lesions and lesion margins, and specimens will undergo HE staining and immunohistochemistry testing. Translated with DeepL.com (free version)

The study is prospective, observational, and diagnostic. Research methodology Equipment Consumables The equipment consumables used include a colonoscopy operating system, pCLE system, fluorescein sodium injection, and single-use endoscopic biopsy forceps. Phase 1 Endoscopy All enrolled person receive fluorescein sodium allergy test before pCLE examination, and fluorescein sodium is injected intravenously after observing no allergic reactions such as dizziness, rash, itching, shock, etc. for 15 min. 30 s later, the pCLE probe is placed from the biopsy clamp channel of enteroscopy, and six different intestinal segments of the cecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum are examined in order and the images are collected, and each intestinal segment of the pCLE of all person is recorded continuously with 3-minute video recordings (operation is kept within 30 min after the procedure). and more than 4000 images are recorded for each person. During the operation, the probe is gently placed on the mucosal surface and kept as perpendicular to the bowel wall as possible, and the movements are gentle, slow, and soft to avoid mucosal injury. The endoscopist is performing targeted biopsies of all suspicious lesions and lesion margins observed in pCLE. Pathological analysis After the specimens are fixed in formalin solution and embedded in paraffin, 2 sections with a thickness of 4 μm are cut. The first section is stained with HE to evaluate the mild inflammation of the intestine; the second section is stained with immunohistochemistry to evaluate the mast cell markers and tight junction proteins in the intestinal wall. Screening of characteristic pCLE images All videos are measured and analyzed by the same experienced endoscopist, after screening all images with good stability, no artifacts, and clear display of microstructures. pCLE images are characterized by a ruler of 10 μm in length in the lower left corner, which allows real-time evaluation of a number of parameters in real time during the operation. Image J image analysis software is used to measure capillary diameter (CD), cell spacing (CS), gland spacing (GS), gland area (GA), and whether sodium fluorescein leak into the crypt lumen in the pCLE field of view. and other indicators. The images are categorized according to the healthy control group and the IBS patient group, and the characteristic images of each group are formed. Phase 2 Diagnostic model construction of IBS patient images Deep learning is performed based on two sets of characteristic images to construct a diagnostic model for IBS. Model evaluation Patients with healthy, IBS, early IBD, or celiac disease (total sample size of 30 cases for total diagnostic model efficacy validation in phases 2 and 3) are examined for pCLE in the same way as in phase 1, and high-quality images are screened for discriminative diagnosis of this model and evaluated for accuracy and inter-evaluator agreement of their diagnostic findings. Phase 3 Construction of a multimodal diagnostic model for images of IBS patients with joint text On the basis of phase 2, the text data of various indicators corresponding to the images are added for joint deep learning to construct a multimodal diagnostic model for IBS patients. Model evaluation Patients with healthy, IBS, early IBD, or celiac disease (total sample size of 30 cases for validation of diagnostic model efficacy in phases 2 and 3 combined) are examined for pCLE in the same way as in phase 1, with high-quality images and various types of indicators screened for discriminatory diagnosis of this model and evaluated for the accuracy of the diagnostic results and inter-appraisal agreement.

No related papers found yet.

None listed