Deep Learning-Based Magnetic Resonance Imaging in Diagnosis and Treatment of Intracranial Aneurysm

Abstract

This study was focused on the positioning of the intracranial aneurysm in the magnetic resonance imaging (MRI) images using the deep learning-based U-Net model, to realize the computer-aided diagnosis of the intracranial aneurysm. First, a network was established based on the three-dimensional (3D) U-Net model, and the collected image data were input into the network to realize the automatic location and segmentation of the aneurysm. The 3D convolutional neural network (CNN) network can extract the aneurysm blood vessels to locate and identify the areas of possible aneurysms. Next, 40 patients highly suspected of intracranial aneurysm were selected as research subjects, and they were subjected to magnetic resonance angiography (MRA) and digital subtraction angiography (DSA) examinations. The results showed that based on the U-Net algorithm model, 40 patients’ hemangiomas were completely contained in the labeling bounding box, one patient’s hemangioma was at the edge of the labeling bounding box, and 4 patients’ hemangiomas were outside the labeling box. The final accuracy coefficient was 88.9%, and it was in good agreement with the doctor’s manual labelling results. Under the 3D CNN network test, the sensitivity, specificity, and accuracy of DSA for intracranial aneurysm were 91.46%, 86.01%, and 90.2%, respectively; the sensitivity, specificity, and accuracy of MRA for intracranial aneurysm were 95.87%, 100%, and 97.19%, respectively. In conclusion, the 3D CNN can successfully realize the positioning of intracranial aneurysm in MRA images, providing a certain theoretical basis for subsequent imaging diagnosis of aneurysm.

Advances in research and application of artificial intelligence and radiomic predictive models based on intracranial aneurysm images

“…The fully connected layers integrate these features to provide a basis for the final decision-making. During the entire training process, backpropagation and optimization algorithms (such as gradient descent) are used to continuously adjust network parameters to minimize prediction errors ( 10 ). CNNs learn discriminative features from input data through convolution, pooling, and activation steps, constructing feature hierarchies from low to high levels ( 2 , 11 ).…”