3D segmentation combining spatial and multi-scale features for intracranial aneurysm.

Abstract

Background: Traditionally, the diagnosis of intracranial aneurysms has relied on the experience of the doctor in assessing the scanning results of radiological imaging technology, which is subjective and inefficient, and it is also limited by the physical strength and energy of the doctor.

Purpose: In order to improve the diagnostic efficiency of doctors and reduce the rate of misdiagnosis and missed diagnosis as much as possible.

Methods: We propose a 3D segmentation network, SMNet, based on the U-Net architecture that combines spatial and multi-scale features. The network can better solve the problem of intracranial aneurysm segmentation on magnetic resonance angiography (MRA) scanning sequences. Specifically, semantic information of different dimensions is extracted at each stage of the encoder by the multi-scale feature extraction block (MSE Block) and the strip volumetric pooling block (SVP Block), respectively. Then, after the fusion of adjacent scale features extracted by the decoder, the weight of features is further redistributed by the quaternary spatial attention block (QSA Block). While focusing on the important features, the ability to discriminate different foregrounds is improved.

Results: Experiments show that the proposed three modules improve the segmentation performance to different degrees. Dice and MIoU have increased by 16.7% and 28% compared to the baseline in the private dataset, and the results of the Aneurysm Detection And segMentation (ADAM) public dataset are 0.482 and 0.389, respectively. It has shown better segmentation quality than 3D medical image segmentation mainstream models.

Conclusion: Our model greatly improves the segmentation results of intracranial aneurysms with MRA images, and makes a certain contribution to the clinical intervention of computer-assisted diagnosis and treatment in this field.