Feasibility Study Of Constructing The Intelligent Assessment System For Image Quality Of Chest Digital Radiography Based On Deep Learning

Part Ⅰ Construction of the Evaluation Index System for Image Quality of Digital Radiography in ChestObjectives The aim of this study was to construct the quantitative evaluation index system for the image quality of chest DR,which can provide an objective reference basis for building an automated chest DR image quality control system.Methods Through literature review,expert consultation,3 rounds of evaluation of 300 chest DR images by 6 radiologists,a knowledge base of image quality of chest DR(IQ-CDR)evaluation indexes including 5 dimensions of layout,pose,gray scale,sharpness and foreign body was formed,and an IQ-CDR evaluation index system consisting of 5 primary indexes and 20 secondary indexes was formed by expert discussion.Ten senior radiologists scored 1138 chest DR on layout,position,gray scale,sharpness and overall quality based on their priori knowledge of image quality,the IQ-CDR evaluation index system was analyzed by measuring the quantitative index parameters of layout,position,gray scale and sharpness using multiple linear regression equations and the correlation between the subjective evaluation of experts,and construct the completed IQ-CDR quantitative evaluation index system.Results A chest DR image quality evaluation index system consisting of 5 primary indicators and 20 secondary indicators was formed.The correlations between the four dimensions of layout,position,gray scale,and sharpness and the overall IQ-CDR were 0.59,0.62,0.95,and 0.94,respectively.The stepwise regression showed that all 10 indicators of layout and position were statistically significant(P<0.05),and multiple linear regression equations were constructed for layout and position,respectively.Conclusions In this study,we constructed a quantitative evaluation index system for image quality in five dimensions:"layout"," position","gray scale","sharpness" and "foreign body"of chest DR images,and the corresponding weight of each index on image quality assessment.Part Ⅱ Developing Intelligent Quality Assessment System for Chest Digital Radiography Based on Deep LearningObjectives This study aims to construct an automated quality assessment system for digital chest radiography based on deep learning algorithms and validate its performance.Methods Based on 1138 chest DR image expert evaluation datasets and 5567 chest DR foreign body annotation datasets,three deep learning models including ResNet-34,CaHDC-RGA and Faster R-CNN convolutional neural network were used to construct IQ-CDR intelligent evaluation model.In the testing phase,the performance of the deep learning model was evaluated using the intra-class correlation coefficient(ICC),Pearson correlation coefficient(r),the mean absolute difference(MAD),mean absolute percentage error(MAPE),and AUC value,sensitivity,and specificity.Results The layout and position models showed high consistency in the measurement of each quantitative index compared to the radiologists(ICC=0.82-0.99,r=0.69-0.99,MAD=0.012.75),and the system showed a small difference in the prediction of scores for both layout and position compared to the experts(MAPE=3.05%,MAPE=5.72%).The gray scale and sharpness models showed high agreement on quality score predictions and binary classification compared to the experts(gray scale:ICC=0.92,r=0.91,MAD=0.45,AUC=0,973;sharpness:ICC=0.90,r=0.89,MAD=0.44,AUC=0.970).The foreign body detection model had a sensitivity of 93.20%,specificity of 92.63%,F1 score of 0.94,and AUC value=0.97.The recall rates within and outside the lung fields were 91.20%and 89.02%.The recall rate was 87.31%and 90.04%for male and female patients.For different age groups,the recall rates were range from 82.67%to 92.67%.Conclusions The automatic image quality assessment system of chest DR consisting of three deep learning models:a)Layout and position model for automatic measurement of quantitative indexes;b)Gray scale and sharpness model for image quality c)Foreign object detection model.The three developed deep learning models all have high performance compared with the manual.