MFDiff: multiscale feature diffusion model for segmentation of 3D intracranial aneurysm from CT images

Intracranial aneurysm is a common life-threatening disease, and the rupture of an intracranial aneurysm carries a high risk of morbidity and mortality. Due to their small size in images, it remains a challenging task to accurately extract the intracranial aneurysms in CT images. In this paper, we propose a multi-scale feature diffusion model, named as MFDiff in short, for segmentation of 3D intracranial aneurysm. The proposed MFDiff includes a feature extraction module and a diffusion model. The feature extraction module is designed to extract features of the original image, and the features act as conditional priors to guide the diffusion model to gradually generate segmentation maps. The diffusion model takes a structure similar to U-Net as backbone, and there is a residual multi-scale feature fusion attention module (RMFA) in the diffusion model, which can adapt to intracranial aneurysms of different size due to multi-scale features. A local CT image dataset is employed for experiment, there are both ruptured and unruptured intracranial aneurysms in the images, and the size of intracranial aneurysms is various, even less than 3 mm. Compared with other popular methods, such as U-Net, GLIA-Net, UNETR++, LinTransUNet, Swin UNETR, the proposed MFDiff shows better performance in intracranial aneurysm segmentation, the segmentation precision is 82.91% when the aneurysms of just size larger than 3 mm are taken into account, and the precision is 75.53% when considering aneurysms of all size.