Application Of Deep Learning Model Based On CTA In Imaging Assessment Of Intracranial Aneurysm

Objective: Two deep learning models for automatic detection of intracranial aneurysms on CTA were established based on Faster RCNN and DenseNet network,and to evaluate their detection performance in detecting intracranial aneurysms.Materials and methods: In this study,2227 patients who underwent CTA examination of craniocerebral artery in the Radiology Department of Affiliated Hospital of Qingdao University from January 2018 to October 2022 were retrospectively collected,including 1527 patients with intracranial aneurysms and 700 patients without aneurysm.DICOM format images of CTA examination were imported into ITK-SNAP software,and aneurysms were sketched manually by a radiologist.After image preprocessing,subjects were randomly divided into a training set and a test set according to a ratio of about 5:2.The Faster RCNN and DenseNet networks were used for training and learning,and two deep learning models for aneurysms detection were developed,and the detection performance of the two models was evaluated respectively.Based on the established DenseNet model,ruptured and unruptured aneurysms in its true positive results were identified.The evaluation parameters included sensitivity,specificity,accuracy,positive predictive value and false positive rate.Differences between groups were compared using chi-square test or fisher exact test.Results:(1)General information: There were 1569 subjects in the training set,including 1069 patients with aneurysms and 500 patients without aneurysms,totaling 1149 aneurysms.A total of 496 aneurysms were detected in 658 subjects in the test set,including 458 patients with aneurysms and 200 patients without aneurysms.(2)Detection performance of the model:(1)The sensitivity,specificity,accuracy,positive predictive value,number of false positives and F1 score of Faster RCNN model for aneurysm detection were 90.32%,83%,80.26%,81.31%,0.16/case and 0.8558.The sensitivity of Faster RCNN model for aneurysms detection with diameter ≥ 5 mm was about 95.16%,the sensitivity for aneurysms detection with diameter < 5 mm was about 83.57%,and the difference was statistically significant(P<0.001).(2)The sensitivity,specificity,accuracy,positive predictive value,number of false positives and F1 score of DenseNet model for detecting aneurysms were 87.90%,76.5%,76.79%,78.7%,0.18/case,and 0.8305,respectively.The sensitivity of Densenet model for detecting aneurysms with diameter≥5 mm was about93.08%,the sensitivity for aneurysms detection with diameter<5 mm was about 80.68%,and the difference was statistically significant(P<0.001).(3)Comparison of detection performance between the two models: The detection performance of Faster RCNN model was better than that of DenseNet model on the whole,and the difference was not statistically significant(all P>0.05).(4)Classification performance of DenseNet model: Among 436 true positive results of DenseNet model,the sensitivity of the model to identify ruptured aneurysms was 53.93%.Conclusions:(1)The automatic intracranial aneurysms detection model based on Faster RCNN and DenseNet has higher detection performance,both as an important auxiliary tool of diagnosing intracranial aneurysms;(2)Faster RCNN model detection performance is better than that of DenseNet model.Objective: Based on the established DenseNet model,this study explored its performance in detecting aneurysms and identifying responsible aneurysms in patients with multiple intracranial aneurysms with subarachnoid hemorrhage,and combined the model with radiologists to evaluate its clinical application value.Materials and methods: In this study,patients with multiple intracranial aneurysms with aneurysmal subarachnoid hemorrhage and 100 patients without aneurysms who underwent intracranial artery CTA examination in the Radiology Department of Affiliated Hospital of Qingdao University from January 2018 to October 2022 were retrospectively collected.After the images of the subjects were randomly sorted,two radiologists read the images blind.Two weeks later,radiologists combined the DenseNet model and reviewed the images again to compare the diagnostic performance before and after the radiologists combined the model.χ2or Fisher exact test was used for statistical analysis of the classified data,and Kruskal-Wallis test was used for statistical analysis of continuous variables.Results:(1)General information: There were 191 aneurysms in 86 patients with multiple intracranial aneurysms with aneurysmal subarachnoid hemorrhage,including 86 ruptured aneurysms and105 unruptured aneurysms.(2)Comparison of the performance of the DenseNet model and the two radiologists in detecting aneurysms and identifying responsible aneurysms:(1)Detection performance: a.The sensitivity of the model was 87.96%,and the average sensitivity of radiologists was86.39%,with no statistically significant difference(P=0.646);b.The specificity,accuracy and positive predictive values of DenseNet model and radiologists were 81%,82%;79.55%,80.19%;80.38%,82.5%,respectively,with no statistically significant difference(all P > 0.05);c.The false positive rate of model and radiologists was 0.22,0.19 / case,and the false negative rate was 0.12,0.14 / case,respectively;(2)Sensitivity of identify responsible aneurysms: the sensitivity of the model was 54.67%,and the average sensitivity of radiologists was 81.58%,with statistically significant difference(P<0.001).(3)Comparison of radiologists’ performance in detecting aneurysms and identifying responsible aneurysms before and after using DenseNet model:(1)Detection performance: a.Sensitivity: the mean sensitivity of radiologists increased from 86.39% to 92.67% after combining the model,with statistically significant difference(P=0.04);b.There was no significant difference in specificity,accuracy and positive predictive value before and after the results of radiologists-binding model(all P > 0.05).c.The false positive rate and false negative rate of radiologists after combining the model were 0.16 and 0.08 / case respectively.(2)Sensitivity of identify responsible aneurysms: radiologists increased from 81.58% to 92.59%,with statistically significant difference(P=0.039).Conclusions:(1)DenseNet model can effectively detect aneurysms in patients with multiple intracranial aneurysms accompanied by subarachnoid hemorrhage,but it is difficult to determine the responsible lesions;(2)Radiologists can improve aneurysms detection result by combining DenseNet model,so as to improve the accuracy of identification responsibility aneurysms.