Study On The Construction And Application Of Deep Learning Based Imaging Aided Diagnosis Model Of Intracranial Aneurysms

ObjectiveTo explore the clinical value of artificial intelligence（AI）computer aided diagnosis（CAD）system based on 3D Universal U-Net(3D U^2-Net)in the detection of intracranial aneurysms（IAs）in 3D-TOF-MRA.Materials and methodsThe MR images of 797 patients with unruptured saccular aneurysms（IAs positive）and150 normal patients without aneurysms（IAs negative）who received physical examination in the Affiliated Hospital of Qingdao University from January 2013 to May 2020 were collected retrospectively.The MR DICOM format images extracted from the PACS system were processed by ITK-SNAP software,and all aneurysms were manually sketched and segmented by a radiologist.After preprocessing MRA images,847 patients were randomly divided into two groups:training group（n=679,including 579 IAs positive and 100 IAs negative）and test group（n=268,including 218 IAs positive and 50 IAs negative）.3D U²-Net general neural network was used to learn and train 67221 MRA images of the training group,and an artificial intelligence CAD system based on deep learning was established.26264 MRA images of the test group were clinically verified.χ2 test or Fisher accurate test were used to compare and analyze the diagnostic efficacy of artificial intelligence CAD system with low seniority radiologist（doctor 1,resident physicians,involved in diagnostic radiology for three years）and senior radiologist（doctor2,attending physicians,involved in radiology diagnosis for for seven years）..Results（1）General data results:1)There were 636 aneurysms in 679 patients in the training group,the diameter range of aneurysms in the training group was 2.0～33.0mm（average size:4.51±3.53mm）;2)There were 235 aneurysms in 268 patients in the test group,and the diameter range of aneurysms in the test group was 2.0～22.0mm（average size 4.71±3.18mm）.There was no significant difference in aneurysm size,location,magnetic field intensity and imaging equipment between the training group and the test group.（2）The detection results of 3D U²-Net model:1)The sensitivity and positive predictive value of 3D U²-Net model are 84.7%and 84.3%,specifity are 88.0%.2)The 3D U²-Net model showed higher sensitivity to>3mm aneurysms than<3mm aneurysms,and the difference was statistically significant（P=0.001<0.05）,the highest sensitivity（96.3%）was detected for aneurysms with diameter>7mm.3)There was no statistically significant difference in the sensitivity of 3D U²-Net model to aneurysm detection in different locations（anterior circulation and posterior circulation）,field intensity（1.5T and 3.0T）,imaging equipment（Philips,Siemens,GE）（P>0.05）.（3）The comparison of the efficacy between doctors and models:1)The sensitivity was doctor 1<3D U²-Net model<doctor 2,and there was no significant difference（doctor1compared with model,P=1.000;doctor 2compared with model,P=0.170）;2)Specificity and positive predictive value were doctor 2>doctor 1>3D U²-Net model.（4）Combining the individual test results of each radiologist with the test results of the3D U²-Net model,the overall diagnostic sensitivity of both radiologists was improved（doctor 1:84.3%to 96.6%;Doctor 2:89.4%to 97.4%）.ConclusionsThe 3D U²-Net model based on deep learning can effectively diagnose intracranial aneurysms with stable performance,Its diagnostic efficacy is roughly similar to that of Chinese doctors in this study,and can be used as an effective tool for the auxiliary diagnosis of aneurysms to some extent.