การจำแนกบริเวณเลือดจากภาพถ่ายกล้องแคปซูล ด้วยวิธีการวิเคราะห์เชิงลึกและคุณสมบัติของสี

Abstract

This engineering project presents the segmentation of bleeding regions in WCE images using a deep learning network and color feature-based. The objective is to assist the diagnosis expert in being more convenient and faster. This research proposed a method to segment bleeding regions from WCE images. It has 3 main sections including: segmentation bleeding regions using Deep Learning network, segmentation bleeding regions using color feature-based and combination of output from Deep Learning network, and color feature-based method. First section has 2 processes: 1. set up training data for high accuracy model 2. segment the bleeding regions. Second section has 3 processes: 1. selecting sample images to make sample color tones 2. making color tones to segment bleeding regions 3. segment the bleeding regions. The last section is the combination of output from 2 sections. The result of section 1 has an initial learn rate of 0.003 and epoch 60, which made a high accuracy model with 87.77 per cent. It has 53.75 percent for loU, 67.06 percent for sensitivity, 91.93 per cent for specificity, and 85.81 per cent for accuracy. Section 2 has 48.77 percent for loU, 69.87 percent for sensitivity, 85.89 percent for specificity, and 81.95 percent for accuracy. The last section has 51.12 percent for loU, 77.30 percent for sensitivity, 83.31 percent for specificity, and 81.83 percent for accuracy. The accuracy of our method in approximately the results is quite good for this project.