Beam selection for radiotherapy

1. A method for determining a radiotherapy treatment plan comprising:

receiving anatomical data for a patient;

generating, via a neural network analyzing the anatomical data, a plurality of fitness values for a plurality of candidate beam orientations, each fitness value corresponding to a candidate beam orientation of the plurality of candidate beam orientations;

determining a selected beam orientation based on the plurality of fitness values;

performing a fluence map optimization (FMO) process on the selected beam orientation; and

determining a dose distribution for the patient based on the FMO process.

2. The method of claim 1 comprising:

configuring a multi-leaf collimator (MLC) to administer the dose distribution at the selected beam orientation.

3. The method of claim 1 , wherein the anatomical data comprises:

at least one image of a tumor within the patient, wherein each of the at least one image comprises contours defining a plurality of organs at risk (OARs); and

a plurality of structure weights associated with the plurality of OARs, wherein each structure weight corresponds to a respective OAR of the plurality of OARs.

4. The method of claim 1 , wherein determining the selected beam orientation comprises using an argmax function to identify a beam orientation associated with a maximum or a minimum fitness value of the plurality of fitness values.

5. The method of claim 1 , wherein determining the selected beam orientation comprises using a tree-search algorithm to identify a solution from a decision space of the neural network.

6. The method of claim 5 , wherein using a tree-search algorithm to identify the solution from the decision space of the neural network comprises iteratively selecting a branch based on a probability distribution of the neural network and updating the probability distribution after each selection.

7. The method of claim 1 , wherein each fitness value of the plurality of fitness values is a prediction of a value based on dual feasibility optimality conditions.

8. The method of claim 1 , wherein the neural network comprises a convolutional up-sampling stage.

9. A method for determining a radiotherapy treatment plan comprising:

receiving anatomical data for a patient;

receiving beam orientation data, wherein the beam orientation data comprises information on pre-selected beam orientations;

generating, via a neural network analyzing the anatomical data and the beam orientation data, a first plurality of fitness values for a plurality of candidate beam orientations, each fitness value corresponding to a candidate beam orientation of the plurality of candidate beam orientations;

determining a first selected beam orientation based on the first plurality of fitness values;

generating updated beam orientation data that comprises information associated with the first selected beam orientation;

generating, via the neural network analyzing the anatomical data and the updated beam orientation data, a second plurality of fitness values for the plurality of candidate beam orientations, each fitness value corresponding to a candidate beam orientation of the plurality of candidate beam orientations;

determining a second selected beam orientation based on the second plurality of fitness values;

performing a fluence map optimization (FMO) process on the first and second selected beam orientations; and

determining a dose distribution for the patient based on the FMO process.

10. The method of claim 9 comprising:

configuring a multi-leaf collimator (MLC) to administer the dose distribution at the first and second selected beam orientations.

11. The method of claim 9 , wherein the anatomical data comprises:

at least one image of a tumor within the patient, wherein each of the at least one image comprises contours defining a plurality of organs at risk (OARs); and

a plurality of structure weights associated with the plurality of OARs, wherein each structural weight corresponds to a respective OAR.

12. The method of claim 9 , wherein determining the first selected beam orientation comprises using an argmax function to identify a beam orientation associated with a maximum or a minimum fitness value of the first plurality of fitness values; and

wherein determining the second selected beam orientation comprises using an argmax function to identify a beam orientation associated with a maximum or a minimum fitness value of the second plurality of fitness values.

13. The method of claim 9 , wherein at least one of determining the first selected beam orientation or the second selected beam orientation comprises using a tree-search algorithm to identify a solution from a decision space of the neural network.

14. The method of claim 13 , wherein using a tree-search algorithm to identify the solution from the decision space of the neural network comprises iteratively selecting a branch based on a probability distribution of the neural network and updating the probability distribution after each selection.

15. The method of claim 9 , wherein each fitness value of the first and second pluralities of fitness values is a prediction of a value based on dual feasibility optimality conditions.

16. The method of claim 9 , wherein generating, via the neural network analyzing the anatomical data and the beam orientation data, the first plurality of fitness values for the plurality of candidate beam orientations comprises:

generating a complement of the beam orientation data; and

generating the first plurality of fitness values for the plurality of candidate beam orientations based on the anatomical data, the beam orientation data, and the complement of the beam orientation data.

17. The method of claim 9 , wherein the neural network comprises a convolutional up-sampling stage.