Using a Convolutional Neural Network and Convolutional Long Short-term Memory to Automatically Detect Aneurysms on 2D-DSA Images (Preprint)

BACKGROUND It is hard to distinguish cerebral aneurysms from overlapping vessels based on 2D DSA images due to their lack of spatial information. OBJECTIVE The aim of this study was to construct a deep learning diagnostic system to improve the ability to detect posterior communicating artery (PCoA) aneurysms on 2D-DSA images and validate the efficiency of the deep learning diagnostic system in 2D-DSA aneurysm detection. METHODS We proposed a two-stage detecting system. First, we established the regional localization stage (RLS) to automatically locate specific detection regions of raw 2D-DSA sequences. Then, in the intracranial aneurysm detection stage (IADS), we constructed the Bi-input+RetinaNet+C-LSTM framework to compare the performance of aneurysm detection with the existing three frameworks. Each of the frameworks had a fivefold cross-validation scheme. The area under the curve (AUC), the receiver operating characteristic (ROC) curve, and mean average precision (mAP) were used to validate the efficiency of different frameworks. The sensitivity, specificity and accuracy were used to identify the abilities of different frameworks. RESULTS A total of 255 patients with PCoA aneurysms and 20 patients without aneurysms were included in this study. The best AUC results of RetinaNet, RetinaNet+C-LSTM, Bi-input+RetinaNet and Bi-input+RetinaNet+C-LSTM were 0.95, 0.96, 0.92 and 0.97, respectively. The sensitivities of RetinaNet, RetinaNet+C-LSTM, Bi-input+RetinaNet, Bi-input+RetinaNet+C-LSTM, and human experts were 89.00% (67.02% to 98.43%), 88.00% (65.76% to 98.06%), 87.00% (64.53% to 97.66%), 89.00% (67.02% to 98.43%), and 90% (68.30% to 98.77%), respectively. The specificity of RetinaNet, RetinaNet+C-LSTM, Bi-input+RetinaNet, Bi-input+RetinaNet+C-LSTM, and human expert were 80.00% (56.34% to 94.27%), 89.00% (67.02% to 98.43%), 86.00% (63.31% to 97.24%), 93.00% (72.30% to 99.56%), and 90% (68.30% to 98.77%), respectively. The accuracies of RetinaNet, RetinaNet+C-LSTM, Bi-input+RetinaNet, Bi-input+RetinaNet+C-LSTM, and human experts were 84.50% (69.57% to 93.97%), 88.50% (74.44% to 96.39%), 86.50% (71.97% to 95.22%), 91.00% (77.63% to 97.72%), and 90.00% (76.34% to 97.21%), respectively. CONCLUSIONS A two-stage aneurysm detection system can reduce the time cost and the computational load. According to our results, more spatial and temporal information can help improve the performances of the frameworks so that Bi-input+RetinaNet+C-LSTM has the best performance compared to the other frameworks. Our study demonstrates that our system can assist doctors in detecting intracranial aneurysms on 2D-DSA images.