Abstract Body: Background
Detection of intracranial aneurysms (IAs) is a time consuming and error prone process. Therefore, solutions that can localize IAs with a high sensitivity are required. Several artificial intelligence (AI)-based automated diagnoses on medical images have recently been reported. We aimed to develop an automated diagnosis system for the location and maximum diameter of IAs on MRI images in the present study.
Methods
In 1310 patients with or without IAs, 937 MRIs were used for training data and 373 cases for test data. The definition of the correct diagnosis of the location was that the center of the aneurysm diagnosed by the AI system is within the area of the IA of ground truth. The nnU-Net was used for the deep learning. Developed AI system was verified with 5-fold cross validation with test data. The maximum diameter was automatically calculated from extracted domes of IAs.
Results
Of the 937 patients used for the model development, 778 (83%) had IAs including 146 patients (19%) with multiple aneurysms and 159 (17%) had no IA. In total 1213 IAs, 78% were small aneurysms of 2 to 5 mm. Of the 373 cases for the validation, 17 (4.5%) had IAs, which is close to the real-world setting. Internal validation of the developed diagnosis model for IA locations showed high efficiency of AUC = 0.92, sensitivity = 15/17 (88%), and false negative rate = 0.44 aneurysms/person. The diameter diagnosis model for the maximum diameter also achieved high accuracy within 1.2 mm of mean absolute error.
Conclusion
We successfully developed a powerful diagnosis model for IAs with AI technique. The present model potentially reduces human effort required for the IA diagnosis.