Enhanced Intracranial Aneurysm Segmentation via Fusion of MedLAM and 3D U-Net Architectures
Shi Ye, Yunfeng Zhou, Jiaqi Wang, Aiping Wu, Mingquan Ye · International Journal of Bio-Inspired Computation · 2025
In response to the inadequate coupling between localisation accuracy and segmentation performance in existing deep learning-based intracranial aneurysm segmentation methods, this study proposes a collaborative segmentation framework utilising the MedLAM architecture and 3D U-Net. Current mainstream approaches typically employ end-to-end single-stage segmentation networks, such as V-Net and nnU-Net, which effectively extract local features. However, given the small volume (usually <5 mm) and highly heterogeneous morphology of intracranial aneurysms in 3D images, segmentation performance frequently diminishes due to the absence of spatial prior constraints. This study introduces an innovative dual mechanism of spatial localisation and feature enhancement through the MedLAM architecture. Firstly, the relative distance regression (RDR) module converts the traditional absolute coordinate-based localisation task into a regression problem of the relative distance field by establishing a voxel-level spatial mapping function, effectively addressing the coordinate offset issue caused by the bending morphology of blood vessels. Secondly, the multi-scale similarity (MSS) mechanism considerably enhances the recognisability of microaneurysms on low-resolution feature maps by dynamically aggregating the similarity responses of feature maps at different levels. Experimental results demonstrate that this method achieves a Dice coefficient of 0.73 on the 500 CTA dataset, marking a 7% improvement over the baseline 3D U-Net.