| Part Ⅰ Construction and comparative study of two automatic detection models for intracranial cystic aneurysms based on computed tomography angiographyObjective:Two deep learning models of intracranial cystic aneurysms,3D Universal U-Net(3D U~2-Net)and Global Localization-based IA Network(GLIA-Net),were constructed respectively with CTA images,and to evaluate the detection performance of intracranial cystic aneurysms.Materials and methods:CTA images of 1571 subjects who underwent head CTA examination in the Affiliated Hospital of Qingdao University from January 2018 to September 2021 were retrospectively collected,including 1271 patients with intracranial cystic aneurysms and 300 patients without aneurysms.DICOM-format CTA images were exported from PACS system and imported into ITK-SNAP software,and all aneurysms were manually sketched and segmented by a radiologist.After preprocessing of CTA images,1571 subjects were randomly divided into training group(n=1034,including 834patients with cystic aneurysm and 200 subjects without aneurysm)and test group(n=537,including 437 patients with cystic aneurysm and 100 subjects without aneurysm,437patients including 100 patients with ruptured cystic aneurysm).3D U~2-Net and GLIA-Net neural network was used to learn and train CTA images of the training group,to construct an automatic detection model for intracranial cystic aneurysm.The trained model was used to test the CTA images of the test group,and the results were compared with the diagnostic criteria to calculate the detection sensitivity,specificity,false positive rate,positive predictive value and F1 score.At the same time,the detection performance of two models for unruptured and ruptured saccular aneurysms were calculated respectively.χ~2 test or Fisher accurate test were used to compare and analyze the diagnostic efficacy between different groups.Results:(1)General characteristics:834 patients in the training group had a total of 918 aneurysms,there were 489 aneurysms in 437 patients in the test group.There was no significant difference in aneurysm size,location and imaging equipment between the training group and the test group.(2)Test results of two models:1)The detection sensitivity,specificity,positive predictive value of 3D U~2-Net model were 85.1%,83.0%and 79.4%,respectively;2)The detection sensitivity,specificity,positive predictive value of GLIA-Net model were 77.5%,89.0%and 84.2%,respectively;3)The sensitivity of the two models for detecting aneurysms with diameter>3 mm was higher than that of aneurysms with diameter<3 mm,and the differences were statistically significant(P<0.001);4)The detection sensitivity of the two models was not affected by aneurysm location,CT imaging equipment and other factors,and the difference were not statistically significant(3D U~2-Net model:P=0.063,0.134;GLIA-Net model:P=0.714,0.509).(3)The comparison of the detection performance between two models:1)The detection sensitivity of 3D U~2-Net model was higher than that of GLIA-Net(P=0.002);2)The detection specificity of 3D U~2-Net model was slightly lower than that of GLIA-Net(P=0.221);3)The F1 scores of both models were higher than 0.80,and the 3D U~2-Net model was slightly higher than that of GLIA-Net model.(4)The detection sensitivity of the two models to unruptured cystic aneurysms were 84.8%and 77.1%,that of ruptured cystic aneurysm were 86.0%and 79.0%,respectively,and the difference were not statistically significant(P=0.770,0.688).Conclusion:The 3D U~2-Net and GLIA-Net model based on deep learning can effectively diagnose intracranial cystic aneurysms and their performance is not affected by aneurysm location,CT imaging equipment and aneurysm rupture bleeding,and the detection performance of 3D U~2-Net model was higher than that of GLIA-Net model.PartⅡComparison of 3D U~2-Net model and radiologists detection performanceObjective:To compare the detection performance of 3D U~2-Net model and radiologists in detecting intracranial cystic aneurysms with CTA images,and to explore the clinical application value of deep learning model in intracranial cystic aneurysm detection.Materials and methods:The data of all subjects included in the study and the detection performance of intracranial cystic aneurysms by 3D U~2-Net model were the same as those in Part I.Besides,before the diagnosis result was known,two radiologists(radiologist 1:3years of experience in imaging diagnosis;radiologist 2:7 years of experience in imaging diagnosis)observed the CTA images of 537 subjects in the test group through PACS platform,and the results were compared with the diagnostic criteria to calculate the detection sensitivity,specificity,false positive rate,positive predictive value and F1 score.After a month washout period,two radiologists observed the CTA images of the test group again with model augmentation,and compared and analyzed the diagnostic performance of radiologists with and without model augmentation.χ~2 test or Fisher accurate test,two-sample T-test were used to compare and analyze the diagnostic efficacy between different groups.Results:(1)The comparison of the efficacy between 3D U~2-Net model and radiologists:1)The detection sensitivity was radiologist 1<3D U~2-Net model<radiologist 2,and there was significant difference between radiologist 1 and model(radiologist 1 compared with model,P=0.004;radiologist 2 compared with model,P=0.308);2)The specificity was radiologist2>radiologist 1>3D U~2-Net model,and there was no significant difference;3)The positive predictive value was radiologist 2>radiologist 1>3D U~2-Net model,and the differences were both statistically significant(radiologist 1 compared with model,P<0.001;radiologist2 compared with model,P<0.001).(2)The 3D U~2-Net model was combined with radiologists’diagnosis:1)The overall detection sensitivity of both radiologists was improved(radiologist 1:77.9%to 90.6%;radiologist 2:87.3%to 92.4%),and the differences were both statistically significant;2)The detection F1 scores of both radiologists were improved,and the improvement degree of radiologist 1 was higher than that of radiologist 2;3)The aneurysm detection time was shortened in both radiologists(radiologist 1:2.47 minutes/case to 1.83 minutes/case;radiologist 2:1.49 minutes/case to 0.74 minutes/case)with model augmentation,and the difference were not statistically significant.Conclusion:The 3D U~2-Net deep learning model based on CTA images can improve the detection performance of radiologists for intracranial cystic aneurysms,which can be used as an effective tool for auxiliary diagnosis of intracranial cystic aneurysms. |
PDF Full Text Request