딥러닝을 사용한 이미지 기반 황달 자가 진단 시스템

Alternative Title
Gijo An
Author(s)
안기조
Alternative Author(s)
Gijo An
Advisor
선우명훈
Department
일반대학원 전자공학과
Publisher
The Graduate School, Ajou University
Publication Year
2020-02
Language
eng
Keyword
Deep Neural NetworkDeep learningJaundiceSelf-diagnosis
Alternative Abstract
The mobile healthcare industry like telemedicine and self-diagnosis are growing with development of IT technology. Jaundice is yellowish pigmentation of the skin and eyes caused by high total bilirubin (T-bilirubin) level in blood due to diseases in the liver, biliary tract and pancreas or a remarkable degradation of their function. Due to measure jaundice which is an important indicator of diseases in these organs, patient must periodically come to the hospital and measure the blood T-bilirubin level through blood collection to trace the changes. In this paper, we proposed image-based jaundice self-diagnostic system using deep learning for those inconvenience and high accuracy jaundice diagnosis. Proposed system consist of a pre-processing unit and a deep learning unit. The pre-processing unit applies color constancy algorithm using patch in image of patient from mobile device, extracts features from segmented sclera area in image. The deep learning unit has 2 stage of deep neural network for high accuracy of estimated T-bilirubin. In first stage, classification network determine whether severe jaundice or not. In next stage, regression network estimate T-bilirubin level. The proposed method was trained and tested using 979 cases of 86 patient from Ajou university hospital with IRB. The test accuracy is 0.93 and AUC is 0.96 in classification network, MAE is 0.0778 in regression network.
URI
https://dspace.ajou.ac.kr/handle/2018.oak/19574
Fulltext

Appears in Collections:
Graduate School of Ajou University > Department of Electronic Engineering > 3. Theses(Master)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Browse