3D deep planning radiotherapy system and method
Systems and methods for three-dimensional dose prediction and treatment planning using a deep learning fully convolutional neural network are disclosed.
1. A computer-implemented method to predict a three-dimensional dose distribution, the method comprising:
receiving, with one or more processors, input data including a three-dimensional voxel image that includes two-dimensional slices, wherein each voxel in the three-dimensional voxel image includes a distance from the voxel to a closest organ surface in three-dimensional space;
providing at least a portion of the input data to a deep fully convolutional neural network (FCNN);
programmatically analyzing in parallel, by the one or more processors, two or more of the two-dimensional slices in the input data;
outputting, by the deep FCNN, a three-dimensional dose distribution prediction that includes a dose prediction for each of the two or more of the two-dimensional slices, wherein the dose prediction for each of the two or more of the two-dimensional slices is independent of dose predictions for other two-dimensional slices and wherein the dose prediction is based on the distance from corresponding voxels to the closest organ surface;
providing, by the one or more processors, the three-dimensional dose distribution prediction as output;
generating, by the one or more processors, a treatment plan based on the three-dimensional dose distribution prediction, wherein the treatment plan includes an electronic file that includes instructions to cause a radiotherapy system to treat a body of a patient associated with the three-dimensional voxel image using radiation; and
transmitting the treatment plan to the radiotherapy system for treatment of the patient associated with the three-dimensional voxel image, wherein the radiotherapy system applies radiation to the patient based on the treatment plan.
2. The method of claim 1 , wherein the distance from the voxel to the closest organ surface in the three-dimensional space is positive if the voxel is within a corresponding organ and wherein the distance is negative if the voxel is outside of the corresponding organ.
3. The method of claim 1 , wherein the deep FCNN is trained based on backpropagation that minimizes a loss function between predicted dose maps and corresponding previously used clinical plans.
4. The method of claim 1 , wherein the treatment plan includes one or more parameters for radiation treatment that include one or more of an amount of radiation to be applied in terms of units of radiation per unit of volume or mass, a time duration for which radiation is to be applied, a count of times radiation is to be applied, a type of radiation to be applied, or a physical location within an organ of the patient.
5. The method of claim 3 , wherein the deep FCNN is further trained by implementing a dropout that reduces overfitting by:
randomly setting one or more neuron in the deep FCNN to zero such that the neuron does not contribute to the backpropagation; and
training remaining neurons in the deep FCNN after the one or more neurons are set to zero.
6. The method of claim 1 , wherein the three-dimensional dose distribution prediction is in units of Gy representing energy deposited per unit of mass (Joule/kilogram).
7. The method of claim 1 , wherein each voxel in the three-dimensional voxel image further includes features channels that describe distances from the voxel to organs of interest that do not receive radiation during application of the radiation to the patient.
8. The method of claim 1 , wherein the radiotherapy system is a linear accelerator (LINAC) and the treatment plan includes one or more of a parameter associated with mechanical movement of one or more components in the LINAC.
9. The method of claim 1 , wherein the radiotherapy system is a linear accelerator (LINAC) and the electronic file is in a data format accepted by the LINAC.
10. The method of claim 9 , wherein the data format includes an extension of a Digital Imaging and Communications in Medicine (DICOM) format including a DICOM-Radiation Therapy (DICOM-RT) data format.
11. The method of claim 10 , wherein the electronic file includes one or more DICOM-RT objects, and wherein the one or more DICOM-RT objects include one or more of an RT image, one or more RT structure sets, an RT plan, an RT dose, an RT beams treatment record, an RT brachy treatment record, or an RT treatment summary.
12. A system comprising:
one or more processors coupled to a computer readable memory having stored thereon software instructions that, when executed by the one or more processors, cause the one or more processors to perform operations including:
receiving input data including a three-dimensional voxel image that includes two-dimensional slices, wherein each voxel in the three-dimensional voxel image includes a distance from the voxel to a closest organ surface in three-dimensional space;
programmatically analyzing in parallel two or more of the two-dimensional slices in the input data
outputting, by a deep fully connected neural network (FCNN), a three-dimensional dose distribution prediction that includes a dose prediction for each of the two or more of the two-dimensional slices, wherein the dose prediction for each of the two or more of the two-dimensional slices is independent of dose predictions for other two-dimensional slices and wherein the dose prediction is based on the distance from corresponding voxels to the closest organ surface;
providing the three-dimensional dose distribution prediction as output;
generating a treatment plan based on the three-dimensional dose distribution prediction, wherein the treatment plan includes an electronic file that includes instructions to cause a radiotherapy system to treat a body of a patient associated with the three-dimensional voxel image using radiation; and
transmitting the treatment plan to the radiotherapy system for treatment of the patient associated with the three-dimensional voxel image, wherein the radiotherapy system applies radiation to the patient based on the treatment plan.
13. The system of claim 12 , wherein:
the distance from the voxel to the closest organ surface in the three-dimensional space is positive if the voxel is within a corresponding organ and wherein the distance is negative if the voxel is outside of the corresponding organ.
14. The system of claim 12 , wherein the one or more processors include a graphics processing unit (GPU) with a plurality of processing units.
15. The system of claim 12 , wherein the one or more processors include a neural network processor.
16. The system of claim 12 , wherein the one or more processors form a portion of an imaging system.
17. The system of claim 12 , wherein each voxel in the three-dimensional voxel image further includes features channels that describe distances from the voxel to organs of interest that do not receive radiation during application of the radiation to the patient.
18. A non-transitory computer readable medium having software instruction stored thereon that, when executed by a processor, cause the processor to perform operations including:
receiving input data including a three-dimensional voxel image that includes two-dimensional slices, wherein each voxel in the three-dimensional voxel image includes a distance from the voxel to a closest organ surface in three-dimensional space;
providing at least a portion of the input data to a deep fully convolutional neural network (FCNN);
programmatically analyzing in parallel two or more of the two-dimensional slices in the input data;
outputting, by the deep FCNN, a three-dimensional dose distribution prediction that includes a dose prediction for each of the two or more of the two-dimensional slices, wherein the dose prediction for each of the two or more of the two-dimensional slices is independent of dose predictions for other two-dimensional slices and wherein the dose prediction is based on the distance from corresponding voxels to the closest organ surface;
providing the three-dimensional dose distribution prediction as output;
generating a treatment plan based on the three-dimensional dose distribution prediction, wherein the treatment plan includes an electronic file that includes instructions to cause a radiotherapy system to treat a body of a patient associated with the three-dimensional voxel image using radiation; and
transmitting the treatment plan to the radiotherapy system for treatment of the patient associated with the three-dimensional voxel image, wherein the radiotherapy system applies radiation to the patient based on the treatment plan.
19. The non-transitory computer readable medium of claim 18 , wherein:
the distance from the voxel to the closest organ surface in the three-dimensional space is positive if the voxel is within a corresponding organ and wherein the distance is negative if the voxel is outside of the corresponding organ.
20. The non-transitory computer readable medium of claim 18 , wherein the deep FCNN is trained based on backpropagation that minimizes a loss function between predicted dose maps and corresponding previously used clinical plans.