Evaluation Of Efficiency And Accuracy Of Lung Nodule Detection Models Based On Deep Convolutional Networks

[Objective]To discuss the significance of clinical applications of deep learning models by comparing pulmonary nodules detection performance of two deep convolutional networks based deep-learning models to that of radiologists with different seniority.[Methods]1.Study subjects:Chest CT images of 225 patients from August 2018 to January 2019 were retrospectively selected with meeting the inclusion and exclusion criteria.2.Reference standards:Two chest imaging experts with more than 20 years of diagnosis experience labeled all the images respectively,identifying and recording the location,size and density of the nodules.With consultation of results from two experts,the reference standard for true positive nodules was confirmed.3.Data collection:Model A and Model B,two kinds of automatic detection models of pulmonary nodules based on deep convolutional networks,respectively interpreted all the cases and recorded the reading time,types and quantities of nodules,along with a junior radiologist with 2-3 years’ working experience and a senior radiologists with more than 10 years’ working experience.With the assist of Model,A,the modified quantities of nodules detected by radiologists with different seniority were recorded as well.4.Statistical methods:Friedman test was applied to compare the reading time among radiologists with different seniority and the two deep-learning models.The systematic difference between the two deep-learning models was revealed through FROC curve analysis.Chi-square test was used to compare the detection rates of pulmonary nodules in location,size and density of radiologists with different seniority and the two deep-learning models,and to compare the changes in the detection rates of pulmonary nodules assisted by the deep learning model.[Results]1.1,047 pulmonary nodules were confirmed in chest CT images of 225 patients,among which 108 nodules distributed in interlobular fissures,13 nodules around the hilus,145 nodules in the central zone,621 nodules in the peripheral zone and 160 in the subpleural areas.In nodule size,there were 572 nodules with diameter<4mm,404 nodules with diameter 4-10mm,and 71 nodules with diameter≥10mm.In nodule density,there were 719 solid nodules,41 subsolid nodules and 287 GGOs.2.The median reading time for each chest CT scan of 225 cases was 175.83 seconds for the junior radiologist,177.02 seconds for the senior radiologist,14.92 seconds for Model A,and 24.81 seconds for Model B.The difference in reading time between junior and senior radiologists was not significant(P=1.0).The reading time of Model A and Model B was significantly shorter than that of manual reading(P<0.05).The difference in reading time between Model A and Model B was significant(P<0.05).3.FROC curve analysis of deep-learning models A and B:With the condition that two false positive nodules were detected in each CT scan,the detection rate of Model A was 94%,and that of Model B was 91.6%.The FROC scores of Model A and Model B were 0.843 and 0.787,respectively.4.The detection rates of 1,047 pulmonary nodules in junior radiologist,senior radiologist,Model A and Model B were 56.92%,70.11%,62.27%and 60.08%,respectively;the difference among groups was significant(P<0.0001).The detection rate of senior radiologist was higher than that of other groups(P<0.05).For nodules with diameter≥4mm,the detection rate of Model A was the highest(P<0.05).5.The detection rates for central.peripheral and subpleural nodules between manual reading and deep-learning models were different(P<0.05).The detection rates for nodules with diameter<4mm and 4-10mm,solid nodules and GGOs were different among four groups(P<0.0001).The detection rates for central and peripheral pulmonary nodules by senior radiologists was higher(P<0.0001,P<0.0001).The detection rates of solid nodules and GGOs were also higher by senior radiologists.For nodules with diameter 4-10mm,the detection rate of Model A was 80.94%(P<0.05).6.Junior and senior radiologists’ detection rates of pulmonary nodules elevated after combination with Model A(P<0.0001,P=0.013),reaching 70.87%and 74.98%respectively.After combination with Model A,junior radiologists’ detection rate was not different with the independent rate of senior radiologists(P=0.707).[Conclusion]Applications of deep-learning model to detect pulmonary nodules can greatly shorten the reading time,and its detection results can assist radiologists to improve the detection rate of pulmonary nodules.