Feasibility Study Of Intelligent Determination Of Lung CT Scan Range Based On Deep Learning Model And Front Scout

Purpose: To investigate the feasibility of accurately predicting the range of chest CT scans based on ortho-localization slices,and to develop a model based on deep learning algorithms to intelligently determine the range of chest CT scans and compare its performance with radiologists.Methods: In this study,clinical data,including demographic data and chest CT image data,were collected from a total of 1984 patients with low-dose CT examinations of the chest from November 2019 to May 2020 at Zhejiang Provincial People’s Hospital.Patients were divided into 12 subgroups according to gender and age（females: 20 to 30,30 to 40,40 to 50,50 to 60,60 to 70,and over 70 years old;males: 20 to 30,30 to 40,40 to 50,50 to 60,60 to 70,and over 70 years old）.A dataset of 600 CT examinations was obtained by randomly selecting 50 individuals from 1984 CT scans by each subgroup,and by using the cross-referencing function of the image archiving and transmission system to detect and record the actual CT lung superior border,the CT lung inferior border,and the anatomical landmarks of the ortho-localization slice（right apical lung,left apical lung,right rib-diaphragm angle,left rib-diaphragm angle）corresponding to the number of axial CT layers were used to explore the relationship between the key points on the localization slice and the actual CT lung borders of the chest.Then,1144 CT data sets were selected from the remaining 1384 CT scans in an8:2 ratio into the training set（80%,n=915）and the validation set（20%,n=229）,and the remaining 240 CT data were used as the test set.Three radiologists annotated 1384 CT localization slices with key points by a professional AI data annotation platform（http://warehouse.healthviewcn.com/）.The annotated training set and validation set images were then preprocessed and data enhanced using Simple ITK（http://www.simpleitk.org）and Python（version 3.6）.The pre-processed and enhanced training and validation set data are used for model construction.The performance of the deep learning algorithm was evaluated in the test set using the actual lung boundaries and the scanning range of the radiographer.Results:（1）There were no significant differences in baseline information including gender and age between the observational study dataset,training set,validation set,and test set（p > 0.05）.（2）The observational study dataset showed that the actual upper lung border exceeded the uppermost lung tip of the orthopositioning slice by 0.77 ± 3.32 mm（95th percentile,5 mm）;the actual lower lung border exceeded the lowermost rib diaphragm angle of the orthopositioning slice by 23.06 ± 17.87 mm（95th percentile,54.9375 mm）.（3）The percentage of marked points between the 3 mm distance threshold range within the observers manually marked on the orthopositioning slice was 100%.The percentage of inter-observer labeled points within the 3mm distance threshold range between two observers was 95.83% to 98.12%.（4）The percentage of correct keypoints predicted by the model for orthostatic localization slices within a distance threshold of 3mm was 95.42% to 100%.（5）The mean difference between the upper and lower lung boundaries predicted by the deep learning model and the actual lung boundaries was 4.72 ± 3.15 mm and 16.50 ±14.06 mm,respectively.the scan range predicted by the model was smaller than that determined manually by the radiographer（298.61 ± 31.24 mm vs.314.79 ± 30.81 mm;P < 0.001）,and the over scan rate was low（29.58% vs.57.08%;2=36.957;P<0.001）.（6）The model-predicted simulated radiation dose for the scan range was lower than that of the radiographer group（0.96±0.4 m Sv vs.1±0.43 m Sv;P<0.001）,with a mean dose reduction of 4.83%.Conclusions:（1）The range of chest CT scans can be accurately predicted using only ortho-positioned slices.（2）The developed model based on deep learning algorithm can accurately identify the anatomical key points of the chest CT ortho-localization slice,and the model can accurately and effectively predict the scan range of chest CT examination.（3）The scan range determined intelligently using the model can reduce the radiation dose compared with that of radiologic technologists.