Abstract | ||
---|---|---|
In this paper, we proposed and validated a novel joint 3D+2D fully convolutional framework for segmenting subcortical structures from magnetic resonance images (MRIs). A 2D Attention U-net (AU-net) following a multi-atlas guided 3D fully convolutional network (MF-net) is constructed in the proposed framework. A novel multi-atlas based encoding block for learning both prior expert information and MRI intensity profile, and a novel attention block for learning structural boundary information are respectively proposed in the 3D MF-net and the 2D AU-net. In the joint 3D+2D framework, the to-be-segmented image and the 2D probability maps for each structure of interest (obtained from the 3D MF-net) were sliced into a 2D image set at each of the three orthogonal views (axial, sagittal, coronal) and then fed into three trained 2D AU-nets, which yields superior segmentation performance. Validation experiments were performed on two datasets respectively contain 16 and 18 T1-weighted MRIs. Compared to several existing state-of-the-art segmentation methods including a multi-atlas joint label fusion method and three representative fully convolutional network methods, the proposed method performed significantly better for a majority of the 12 subcortical structures, with the overall mean Dice scores being respective 0.917 and 0.865 for the two datasets. |
Year | DOI | Venue |
---|---|---|
2019 | 10.1007/978-3-030-32248-9_34 | Lecture Notes in Computer Science |
Keywords | DocType | Volume |
Subcortical structures,Segmentation,Multi-atlas,Attention U-net,Joint fully convolutional framework | Conference | 11766 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Jiong Wu | 1 | 2 | 3.06 |
Yue Zhang | 2 | 0 | 1.69 |
Xiaoying Tang | 3 | 8 | 8.79 |