Bottom Double Branch Path Networks With Confidence Calibration for Intracranial Aneurysms Detection in <scp>3D MRA</scp>

ABSTRACT Intracranial aneurysms (IAs) are characterized by abnormal dilation of the brain blood vessel wall, the rupture of which often leads to subarachnoid hemorrhage with a high mortality rate. Current detections rely heavily on radiologists' interpretation of magnetic resonance angiography (MRA) images, but manual identification is time‐consuming and laborious. Therefore, it is urgent to carry out automatic detection tools for IAs, and various intelligent models have been developed in recent years. However, the size of IAs is relatively small compared with the high voxel resolution MRA images, and thus the data imbalance leads to a high false positive (FP) rate. To address these challenges, we have proposed an innovative 3D voxel detection framework based on Feature Pyramid Network (FPN) architecture, which is called bottom double branch path network with confidence calibration (BCOC for short). BCOC shows better effects on small objects for preserving diversities of feature maps and also creates efficient feature extractors by reducing the number of channels per layer, making it particularly advantageous for handling large three‐dimensional resolutions. Additionally, optimal transport (OT) has been applied for matching the detection and ground truth bounding boxes during the post‐process phase to refine bounding box positions, thereby further improving the detection performance. Moreover, the confidence score of model output is calibrated via calibration loss during training to make correct detections with higher confidence and wrong detections with lower confidence, which can reduce the FP rate. Our proposed model achieves mean average precision (AP) of 0.8186 and 0.8533, sensitivity of 93.91% and 98.43%, FPs/case of 0.1332 and 0.0541 on two public MRA datasets including cases with IAs collected from different hospitals, respectively, outperforming other state‐of‐the‐art methods. The results show that BCOC is a promising detection method for IAs automatic recognition.