Fig 1.
Slices from three different viewpoints.
Pancreas is highlighted in red and is a small organ with vague boundary and irregular shape.
Fig 2.
Illustration of the segmentation results of two different input regions.
A smaller input region improves the segmentation performance.
Fig 3.
Illustration of the coarse-to-fine framework.
The framework segments the pancreas through coronal, sagittal and axial viewpoints and fuses the results into a new segmentation volume.
Fig 4.
Illustration of the proposed multi-scale attention net.
Dense block is show in Fig 5 and the proposed attention module is shown in Fig 6.
Fig 5.
Illustration of the dense block.
Dense block is composed of 6 convolution layers to obtain a large receptive field and skip connections are used to avoid the risk of network degradation, gradient vanish and gradient explosion.
Fig 6.
Illustration of the proposed attention module.
It is composed of a spatial attention module and a channel attention module, and make the network focus on the most relevant regions and channels.
Table 1.
Evaluation for coarse segmentation of three axes and fusion segmentation is reported in the table.
Table 2.
Comparison between our approach and the state-of-the-art approaches on NIH pancreas dataset.
Fig 7.
An example of the segmentation results.
Compared with the coarse segmentation, fine segmentation result has been improved from 70.45% to 84.31%. However, we can observe that we cannot predict the boundary of the pancreas very well.
Table 3.
Ablation study of our proposed 2.5D U-net and proposed attention module.
AG: The attention gate proposed in [24]. DA: The dual attention module proposed in [22]. HA: Our proposed attention module. Params: The number of parameters.
Table 4.
Analysis of the DSC distribution of all slices.