Using radiation dose information for automatic organ segmentation model training

1. A method for training a machine learning model for automatic organ segmentation, the method comprising:

receiving, by a processor, an image of one or more pre-contoured organs, the image comprising a plurality of voxels;

executing, by the processor, a machine learning model using the image to output predicted organ labels for the plurality of voxels of the image;

determining, by the processor, differences between corresponding predicted organ labels and expected organ labels for the plurality of voxels of at least one pre-contoured organ;

determining, by the processor, radiation dose levels that correspond to the plurality of voxels of the image;

determining, by the processor, weights for the plurality of voxels based on the radiation dose levels of the respective voxels; and

training, by the processor, the machine learning model based on the differences and the weights for the plurality of voxels.

2. The method of claim 1 , wherein determining a weight for a voxel of the plurality of voxels comprises:

comparing, by the processor, a radiation dose level for the voxel to a threshold; and

determining, by the processor, the weight for the voxel to be zero responsive to determining the radiation dose level does not exceed the threshold.

3. The method of claim 1 , wherein determining the radiation dose levels for the plurality of voxels comprises:

receiving, by the processor, a radiation therapy treatment plan comprising treatment attributes for a patient, the treatment attributes comprising radiation dose levels for the plurality of voxels; and

identifying, by the processor, the radiation dose levels from the radiation therapy treatment plan.

4. The method of claim 1 , wherein determining the radiation dose levels for the plurality of voxels comprises:

receiving, by the processor, a radiation therapy treatment plan comprising treatment attributes for a patient, the treatment attributes comprising field geometry settings for a radiation therapy machine treating the patient; and

determining, by the processor, the radiation dose levels based on the field geometry settings.

5. The method of claim 4 , wherein determining the radiation dose levels based on the field geometry settings comprises:

determining, by the processor, whether the plurality of voxels of the image correspond to an area on a body of the patient at which the radiation therapy machine directed radiation;

wherein determining, by the processor, the weights for the plurality of voxels comprises determining, by the processor, binary weights for the identified voxels based on the determination of whether the plurality of voxels correspond to an area on a body of the patient at which the radiation therapy machine directed radiation.

6. The method of claim 1 , wherein determining a weight for a voxel of the plurality of voxels comprises:

determining, by the processor, a difference between a radiation dose level that corresponds to the voxel and a target dose level; and

determining, by the processor, the weight by applying a sigmoid function to the difference.

7. The method of claim 6 , wherein determining the weight for the voxel further comprises:

adjusting, by the processor, the difference between the radiation dose level and the target dose level based on a scaling factor,

wherein applying the sigmoid function to the difference comprises applying, by the processor, the sigmoid function to the adjusted difference.

8. The method of claim 1 , wherein training the machine learning model comprises:

assigning, by the processor, the determined weights to corresponding squared differences between corresponding predicted organ labels and expected organ labels for the plurality of voxels.

9. The method of claim 1 , further comprising:

receiving, by the processor, a second image;

executing, by the processor, the trained machine learning model using the second image to obtain an output of contoured organs; and

adjusting, by the processor, a field geometry setting of a radiotherapy machine based on the output of contoured organs.

10. The method of claim 9 , wherein receiving the second image comprises receiving a three-dimensional image;

wherein executing the trained machine learning model using the second image comprises applying, by the processor, individual voxels of the image to the trained machine learning model, and

wherein obtaining the output of the contoured organs comprises obtaining a three-dimensional model of the contoured organs, the three-dimensional model comprising predicted labels for voxels of the three-dimensional model indicating contoured organs that correspond to the respective voxels.

11. A system for training a machine learning model for automatic organ segmentation, the system comprising:

a processor configured to execute instructions stored on a non-transitory computer-readable medium to:

receive an image of one or more pre-contoured organs, the image comprising a plurality of voxels;

execute a machine learning model using the image to output predicted organ labels for the plurality of voxels of the image;

determine differences between corresponding predicted organ labels and expected organ labels for the plurality of voxels of at least one pre-contoured organ;

determine radiation dose levels that correspond to the plurality of voxels of the image;

determine weights for the plurality of voxels based on the radiation dose levels of the respective voxels; and

train the machine learning model based on the differences and the weights for the plurality of voxels.

12. The system of claim 11 , wherein the processor is configured to determine a weight for a voxel of the plurality of voxels by:

comparing a radiation dose level for the voxel to a threshold; and

determining the weight for the voxel to be zero responsive to determining the radiation dose level does not exceed the threshold.

13. The system of claim 11 , wherein the processor is configured to determine the radiation dose levels for the plurality of voxels by:

receiving a radiation therapy treatment plan comprising treatment attributes for a patient, the treatment attributes comprising radiation dose levels for the plurality of voxels; and

identifying the radiation dose levels from the radiation therapy treatment plan.

14. The system of claim 11 , wherein the processor is configured to determine the radiation dose levels for the plurality of voxels by:

receiving a radiation therapy treatment plan comprising treatment attributes for a patient, the treatment attributes comprising field geometry settings for a radiation therapy machine treating the patient; and

determining the radiation dose levels based on the field geometry settings.

15. The system of claim 14 , wherein the processor is configured to determine the radiation dose levels based on the field geometry settings by:

determining whether the plurality of voxels of the image correspond to an area on a body of the patient at which the radiation therapy machine directed radiation; and

determining binary weights for the identified voxels based on the determination of whether the plurality of voxels correspond to an area on a body of the patient at which the radiation therapy machine directed radiation.

16. The system of claim 11 , wherein the processor is configured to determine a weight for a voxel of the plurality of voxels by:

determining a difference between a radiation dose level that corresponds to the voxel and a target dose level; and

determining the weight by applying a sigmoid function to the difference.

17. The system of claim 16 , wherein the processor is configured to determine the weight for the voxel by:

adjusting the difference between the radiation dose level and the target dose level based on a scaling factor,

wherein the processor is configured to apply the sigmoid function to the difference by applying the sigmoid function to the adjusted difference.

18. The system of claim 11 , wherein the processor is configured to train the machine learning model by:

assigning the determined weights to corresponding squared differences between corresponding predicted organ labels and expected organ labels for the plurality of voxels.

19. The system of claim 11 , wherein the processor is configured to:

receive a second image;

execute the trained machine learning model using the second image to obtain an output of contoured organs; and

adjust a field geometry setting of a radiotherapy machine based on the output of contoured organs.

20. The system of claim 19 , wherein the processor is configured to receive the second image by receiving a three-dimensional image;

wherein the processor is configured to execute the trained machine learning model using the second image by applying individual voxels of the image to the trained machine learning model, and

wherein the processor is configured to obtain the output of the contoured organs by obtaining a three-dimensional model of the contoured organs, the three-dimensional model comprising predicted labels for voxels of the three-dimensional model indicating contoured organs that correspond to the respective voxels.