Two and a half dimensional convolutional neural network for predicting hematoma expansion in non-contrast head computerized tomography images

1. A method comprising:

receiving a three dimensional (3D) non-contrast computed tomography (NCCT) image of a head that includes at least one intracranial hemorrhage;

generating a plurality of two dimensional (2D) NCCT images of the head using the 3D NCCT image of the head;

generate, by applying a 2D U-Net convolutional neural network (CNN) on each of the plurality of 2D NCCT images, hematoma masks for the plurality of 2D NCCT hematoma images;

determining a bounding box that contains a hematoma from the hematoma masks such that a center of the hematoma in each of the plurality of 2D NCCT hematoma images coincides,

determining a fixed-size 3D region of interest (ROI) containing the hematoma from the bounding box;

extracting the fixed-size 3D ROI from the 3D NCCT image;

generating a plurality of 2D NCCT slices of the fixed-size 3D ROI that is extracted;

predicting a hematoma expansion (HE) of the at least one intracranial hemorrhage by inputting each of the plurality of 2D NCCT slices to a 2D NASNet CNN, each of the plurality of 2D NCCT slices being a feature channel of the 2D NASNet CNN, and combining a plurality of outputs of the 2D NASNet CNN to generate an HE prediction score of the at least one intracranial hemorrhage; and

providing the HE prediction score.

2. The method of claim 1 , wherein the size of the fixed-size 3D ROI contains all hematomas in the 3D NCCT image of the head.

3. The method of claim 1 , wherein determining the bounding box comprises:

identifying a volume of the hematomas based on the plurality of 2D NCCT hematoma images; and

determining a size of the fixed-size 3D ROI based on the volume.

4. The method of claim 1 , wherein generating the hematoma masks comprises:

identifying, by applying the 2D U-Net CNN, the hematoma based on the plurality of 2D NCCT images of the head; and

generating the plurality of 2D NCCT hematoma images in which the hematoma is segmented based on the plurality of 2D NCCT images of the head.

5. The method of claim 1 , wherein the plurality of 2D NCCT images of the head corresponds to an entire image of the head.

6. The method of claim 5 , wherein the fixed-size 3D ROI corresponds to less than the entire image of the head.

7. A device comprising:

at least one memory configured to store program code;

at least one processor configured to read the program code and operate as instructed by the program code, the program code including:

receiving code configured to cause the at least one processor to receive a three dimensional (3D) non-contrast computed tomography (NCCT) image of a head;

generating code configured to cause the at least one processor to generate a plurality of two dimensional (2D) NCCT images of the head using the 3D NCCT image of the head;

mask generating code configured to cause the at least one processor to generate, by applying a 2D U-Net convolutional neural network (CNN) on each of the plurality of 2D NCCT images, hematoma masks for the plurality of 2D NCCT hematoma images;

bounding box code configured to cause the at least one processor to determine a bounding box that contains a hematoma from the hematoma masks such that a center of the hematoma in each of the plurality of 2D NCCT hematoma images coincides;

determining code configured to cause the at least one processor to determine a fixed-size 3D region of interest (ROI) containing the hematoma from the bounding box;

extracting code configured to cause the at least one processor to extract the fixed-size 3D ROI from the 3D NCCT image;

slicing code configured to cause the at least one processor to generate a plurality of 2D NCCT slices of the fixed-size 3D ROI that is extracted;

HE prediction code configured to cause the at least one processor to predict a hematoma expansion (HE) of the at least one intracranial hemorrhage by inputting each of the plurality of 2D NCCT slices to a 2D NASNet CNN, each of the plurality of 2D NCCT slices being a feature channel of the 2D NASNet CNN, and combining a plurality of outputs of the 2D NASNet CNN to generate an HE prediction score of the at least one intracranial hemorrhage; and

providing code configured to cause the at least one processor to provide the HE prediction score.

8. The device of claim 7 , wherein the bounding box code comprises:

identifying code configured to cause the at least one processor to identify a volume of the hematoma based on the plurality of 2D NCCT hematoma images; and

second determining code configured to cause the at least one processor to determine a size of the fixed-size 3D ROI based on the volume.

9. The device of claim 7 , wherein the mask generating code comprises:

identifying code configured to cause the at least one processor to identify, by applying the 2D U-Net CNN, the hematoma based on the plurality of 2D NCCT images of the head; and

second generating code configured to cause the at least one processor to generate the plurality of 2D NCCT hematoma images in which the hematoma is segmented based on the plurality of 2D NCCT images of the head.

10. The device of claim 7 , wherein the plurality of 2D NCCT images of the head corresponds to an entire image of the head.

11. The device of claim 10 , wherein the fixed-size 3D ROI corresponds to less than the entire image of the head.

12. A non-transitory computer-readable medium storing computer code that, when executed by one or more processors, cause the one or more processors to at least:

receive a three dimensional (3D) non-contrast computed tomography (NCCT) image of a head;

generate a plurality of two dimensional (2D) NCCT images of the head using the 3D NCCT image of the head;

generate, by applying a 2D U-Net convolutional neural network (CNN) on each of the plurality of 2D NCCT images, hematoma masks for the plurality of 2D NCCT hematoma images;

determine a bounding box that contains a hematoma from the hematoma masks such that a center of the hematoma in each of the plurality of 2D NCCT hematoma images coincides;

determine a fixed-size 3D region of interest (ROI) containing the hematoma from the bounding box;

extracting the fixed-size 3D ROI from the 3D NCCT image;

generating a plurality of 2D NCCT slices of the fixed-size 3D ROI that is extracted;

predicting a hematoma expansion (HE) of the at least one intracranial hemorrhage by inputting each of the plurality of 2D NCCT slices to a 2D NASNet CNN, each of the plurality of 2D NCCT slices being a feature channel of the 2D NASNet CNN, and combining a plurality of outputs of the 2D NASNet CNN to generate an HE prediction score of the at least one intracranial hemorrhage; and

provide the HE prediction score.

13. The non-transitory computer-readable medium of claim 12 , wherein to generate the hematoma masks, the computer code is further configured to cause the one or more processors to:

identify, by applying the first 2D CNN, the hematoma based on the plurality of 2D NCCT images of the head; and

generate the plurality of 2D NCCT hematoma images in which the hematoma is segmented based on the plurality of 2D NCCT images of the head.

14. The non-transitory computer-readable medium of claim 12 , wherein the one or more processors determine the bounding box by at least:

identifying a volume of the hematomas based on the plurality of 2D NCCT hematoma images; and

determining a size of the fixed-size 3D ROI based on the volume.

15. The non-transitory computer-readable medium of claim 12 , wherein the plurality of 2D NCCT images of the head corresponds to an entire image of the head.

16. The non-transitory computer-readable medium of claim 15 , wherein the fixed-size 3D ROI corresponds to less than the entire image of the head.