OTO-Net: An Automated MRA Image Segmentation Network for Intracranial Aneurysms

Abstract

Intracranial aneurysms are local dilations of the cerebral blood vessels; people with intracranial aneurysms have a high risk to cause bleeding in the brain, which is related to high mortality and morbidity rates. Accurate detection and segmentation of intracranial aneurysms from Magnetic Resonance Angiography (MRA) images are essential in the clinical routine. Manual annotations used to assess the intracranial aneurysms on MRA images are substantial interobserver variability for both aneurysm detection and assessment of aneurysm size and growth. Many prior automated segmentation works have focused their efforts on tackling the problem, but there is still room for performance improvement due to the significant variability of lesions in the location, size, structure, and morphological appearance. To address these challenges, we propose a novel One-Two-One Fully Convolutional Networks (OTO-Net) for intracranial aneurysms automated segmentation in MRA images. The OTO-Net uses full convolution to achieve intracranial aneurysms automated segmentation through the combination of downsampling, upsampling, and skip connection. In addition, loss ensemble is used as the objective function to steadily improve the backpropagation efficiency of the network structure during the training process. We evaluated the proposed OTO-Net on one public benchmark dataset and one private dataset. Our proposed model can achieve the automated segmentation accuracy with 98.37% and 97.86%, average surface distances with 1.081 and 0.753, dice similarity coefficients with 0.9721 and 0.9813, and Hausdorff distance with 0.578 and 0.642 on these two datasets, respectively.

Staged cluster transformers for intracranial aneurysms segmentation from structure fused 3D MRA

“…In HD calculations, the maximum distance quantile is set to 95% (HD95) in order to eliminate the little distance caused by some outliers 5 …”

“…A network 44 was proposed for IAs segmentation combining 3D U-Net, residual connection and attention gate. Ye et al 5 proposed a OTO-Net with three consecutive encoding and decoding structures to segment IAs on MRA. Despite some performance improvements gained by these models, they suffer from the local receptive field and the resulting missed detection of IAs, as has been said before.…”

“…The current conventional imaging-based examination methods for IAs include Digital Subtracted Angiography (DSA), Computed Tomography Angiography (CTA), and Magnetic Resonance Angiography (MRA), among which MRA is generally the first choice for patients because it is a noninvasive method with simpler operation and higher safety that does not require radiation exposure. 5 At present, radiologists use MRA images to perform 3D visualization and manually sketch the location of IAs by experience, which is highly subjective, labor intensive and time consuming. In view of this, some automatic segmentation algorithms for IAs have gradually appeared driven by the development of artificial intelligence.…”