A study on deep learning-based smartphone application for scleral jaundice diagnosis
공막 황달 진단을 위한 딥러닝 기반 스마트폰 애플리케이션 연구
초록/요약
Jaundice is a symptom of yellow pigmentation on the sclera and skin due to the deposition of bilirubin. Therefore, jaundice is a prognostic symptom that can diagnose various diseases at an early stage, and is important information to know the treatment progress of the patient. There are also areas that are difficult to detect with the naked eye, even when abnormal bilirubin levels occur. This study develops a patient-friendly jaundice diagnosis smartphone application. The proposed deep-learning-based smartphone application can detect 1.5 to 3.0 mg/dl TSB, which is easily missed by the naked eye. Patients can use this application to ensure that they do not miss an appropriate visit in hospital. The system predicts serum total bilirubin levels using deep learning-based regression analysis of scleral photos taken with a smartphone's built-in camera. A color correction patch is used together to achieve consistent performance in various lighting environments and to segment only the patient's eye area. The region segmented by the color correction patch provides predicted bilirubin levels to the patient through deep learning-based regression analysis. Patients were randomly seperated to a training or a test cohort. Intra-class correlation coefficient values for predicted TSB derived from images taken repeatedly at the same visit for the same patient showed good confidence (0.86). Strong correlation between the measured TSB and the predicted TSB was observed from 1.5 mg/dl or higher at the onset of scleral discoloration (Spearman rank order correlation=0.70, P<0.01). The ROC curve for predicted TSB showed an area under the curve of 0.93, indicating good performance as a predictor of hyperbilirubinemia (P<0.001). With cut-off TSB 1.5 mg/dl, the predictive sensitivity of hyperbilirubinemia was 80.0% and specificity was 92.6%. This study present an estimated total serum bilirubin level that will require a hospital visit when using the application. Therefore, the proposed application is effective as an early diagnosis method for patients who need regular hospital visits for chronic liver, biliary tract and pancreatic diseases or for jaundice diagnosis.
more목차
I. Introduction 1
A. Background 1
B. Purpose of the study 6
C. Dissertion outline 8
II. Related Works 9
A. Convolution Neural Network 9
1. LeNet-5 12
2. VGG-16 13
B. Jaundice self-diagnosis using smartphone 17
1. BiliCam 17
2. Biliscreen 19
3. Jaundice Eye Color Index 21
4. Previous Research 24
III. Patients and Dataset Chracteristic 26
A. Patients 26
B. Baseline Characteristic 29
IV. Proposed Jaundice Diagnostic Application 31
A. Pre-processing 33
B. Deep Learning Architecture 34
V. Results and Validation 40
VI. Conclusion 54
Bibliography 58
