Apparatus and method for image segmentation
View Patent ↗An apparatus for image segmentation according to an embodiment includes an acquirer configured to acquire one or more images in which an object is photographed and a segmentation performer configured to perform segmentation on the one or more images using a segmentation model which is deep learned through a plurality of images, in which the segmentation model is a U-Net-based model including a first type module based on depth-wise separable convolution (DSC) and a second type module based on global context network (GCNet).
1. An apparatus for image segmentation, which is implemented with a computing device that includes one or more processors and a memory for storing one or more programs executed by the one or more processors, the apparatus comprising:
an acquirer configured to acquire one or more images in which an object is photographed; and
a segmentation performer configured to perform segmentation on the one or more images using a segmentation model which is deep learned through a plurality of images,
wherein the segmentation model is a U-Net-based model including a first type module based on depth-wise separable convolution (DSC) and a second type module based on global context network (GCNet),
wherein the first type module is configured to include a plurality of depth-wise convolution layer blocks for extracting feature information of a feature map and a plurality of point-wise convolution layer blocks including a first point-wise convolution layer block and a second point-wise convolution layer block for controlling a number of channels of the feature map,
wherein the first type module is configured to:
calculate a map for extracting feature information by repeatedly applying the depth-wise convolution layer block and the point-wise convolution layer block to an input feature map, and increase a number of output channels by using the first point-wise convolution layer block and then adjust the increased number of the output channels to the number before increasing by using the second point-wise convolution layer block;
calculate a map for controlling the number of channels by applying only the point-wise convolution layer blocks to the input feature map; and
sum the map for extracting the feature information and the map for adjusting the number of channels and output a result of the summation.
2. The apparatus of claim 1 , wherein the one or more images include a tomographic image of a brain obtained by magnetic resonance imaging (MRI).
3. The apparatus of claim 2 , wherein the segmentation performer is configured to segment the tomographic image of the brain into a plurality of sections, and determine one or more sections satisfying a preset condition among the plurality of sections as a stroke lesion.
4. The apparatus of claim 1 , wherein the segmentation model is a U-Net-based model in which at least some of a plurality of convolution layer blocks in the segmentation model are replaced with the first type module.
5. The apparatus of claim 1 , wherein the segmentation model is a U-Net-based model in which the second type module is disposed between an encoder in which down sampling is performed and a decoder in which up sampling is performed in the segmentation model.
6. The apparatus of claim 1 , wherein the second type module is configured to include a first convolution layer block for extracting feature information of a feature map, a second convolution layer block, and a global context block (GCBlock) based on the global context network.
7. The apparatus of claim 6 , wherein the second type module is configured to:
calculate a global feature map by applying the first convolution layer block, the second convolution layer block, and the global context block to an input feature map; and
sum the input feature map and the global feature map and output a result of the summation.
8. A method for image segmentation performed by a computing device that includes one or more processors and a memory for storing one or more programs executed by the one or more processors, the method comprising:
acquiring one or more images in which an object is photographed; and
performing segmentation on the one or more images using a segmentation model which is deep learned through a plurality of images,
wherein the segmentation model is a U-Net-based model including a first type module based on depth-wise separable convolution (DSC) and a second type module based on global context network (GCNet),
wherein the first type module is configured to include a plurality of depth-wise convolution layer blocks for extracting feature information of a feature map and a plurality of point-wise convolution layer blocks including a first point-wise convolution layer block and a second point-wise convolution layer block for controlling the number of channels of the feature map,
wherein the first type module is configured to:
calculate a map for extracting feature information by repeatedly applying the depth-wise convolution layer block and the point-wise convolution layer block to an input feature map, and, in calculating the map for extracting feature information, increase a number of output channels using a first point-wise convolution layer block of the plurality of point-wise convolution layer blocks and then adjust the increased number of the output channels to the number before increasing by using a second point-wise convolution layer block;
calculate a map for controlling the number of channels by applying only the point-wise convolution layer block to the input feature map; and
sum the map for extracting the feature information and the map for adjusting the number of channels and output a result of the summation.
9. The method of claim 8 , wherein the one or more images include a tomographic image of a brain obtained by magnetic resonance imaging (MRI).
10. The method of claim 9 , wherein in the performing of segmentation, the tomographic image of the brain is segmented into a plurality of sections, and one or more sections satisfying a preset condition among the plurality of sections is determined as a stroke lesion.
11. The method of claim 8 , wherein the segmentation model is a U-Net-based model in which at least some of a plurality of convolution layer blocks in the segmentation model are replaced with the first type module.
12. The method of claim 8 , wherein the segmentation model is a U-Net-based model in which the second type module is disposed between an encoder in which down sampling is performed and a decoder in which up sampling is performed in the segmentation model.
13. The method of claim 8 , wherein the second type module is configured to include a first convolution layer block for extracting feature information of a feature map, a second convolution layer block, and a global context block (GCBlock) based on the global context network.
14. The method of claim 13 , wherein the second type module is configured to:
calculate a global feature map by applying the first convolution layer block, the second convolution layer block, and the global context block to an input feature map; and
sum the input feature map and the global feature map and output a result of the summation.