Deep reinforcement learning for computer assisted reading and analysis
Methods and systems of revising image models for nuclear imaging are disclosed. A system receives first patient scan data corresponding to one or more nuclear imaging scans performed on one or more individuals and generates a first reconstructed image by applying a first image model to the first patient scan data. Feedback is received regarding the first reconstructed image and the feedback is provided as an input to a deep-reinforcement learning process. The deep-reinforcement learning process is configured to generate at least one modification for the first image model based on the feedback regarding the first reconstructed image. A second image model is generated by applying the at least one modification generated by the deep-reinforcement learning process to the first image model.
1. A method of revising image models for nuclear imaging, comprising:
receiving, by a system, first patient scan data corresponding to a nuclear imaging scan of an individual;
receiving, by the system, preference data associated with the first patient scan data;
selecting, by the system, a first image model from a plurality of image models based on at least a first portion of the preference data, and modifying, by the system, the first image model based on at least a second portion of the preference data;
generating, by the system, a first reconstructed image by applying the first image model to an entirety of the first patient scan data, the first reconstructed image comprising an imaging target at a first location;
receiving, by the system, a feedback regarding the first reconstructed image;
generating weighted feedback based on applying a first weight to a first portion of the feedback and a second weight to a second portion of the feedback;
providing, by the system, the weighted feedback as an input to a deep-reinforcement learning process, wherein the deep-reinforcement learning process is configured to extract features from the weighted feedback, and generate at least one modification for the first image model based on the extracted features of the weighted feedback regarding the first reconstructed image;
generating, by the system, a second image model by applying the at least one modification generated by the deep-reinforcement learning process to the first image model;
generating, by the system, a second reconstructed image by applying the second image model to the entirety of the first patient scan data, the second reconstructed image comprising the imaging target at a second location that is different from the first location.
2. The method of claim 1 , wherein the first image model is a predefined image model.
3. The method of claim 1 , wherein the feedback comprises system generated feedback or expert generated feedback.
4. The method of claim 3 , wherein the system generated feedback comprises an artifact reduction, an acquisition protocol, or a reconstruction protocol.
5. The method of claim 3 , wherein the expert generated feedback comprises a score indicative of an acceptability of the first reconstructed image or the at least one modification.
6. The method of claim 1 , wherein the feedback comprises a modified reconstructed image.
7. The method of claim 1 , wherein the deep-reinforcement learning process comprises a convolution neural network or a text evaluation network.
8. The method of claim 1 , comprising:
receiving, by the system, second patient scan data corresponding to one or more additional nuclear imaging scans; and
generating, by the system, a third reconstructed image by applying the second image model to the second patient scan data.
9. The method of claim 1 , wherein the preference data comprises machine-learning training sets.
10. The method of claim 9 , wherein the method comprises training the deep-reinforcement learning process based on the machine-learning training sets, wherein the trained deep-reinforcement learning process is configured to extract the features from the weighted feedback.
11. A system, comprising:
an imaging modality configured to generate patient scan data in response to a nuclear imaging scan of an individual during a first imaging period; and
a computer configured to:
receive the patient scan data generated during the first imaging period;
receive preference data associated with the first patient scan data;
select a first image model from a plurality of image models based on at least a first portion of the preference data, and modify the first image model based on at least a second portion of the preference data;
generate a first reconstructed image by applying the first image model to an entirety of the patient scan data, the first reconstructed image comprising an imaging target at a first location;
receive a feedback regarding the first reconstructed image;
generate weighted feedback based on applying a first weight to a first portion of the feedback data and a second weight to a second portion of the feedback data;
provide the weighted feedback as an input to a deep-reinforcement learning process, wherein the deep-reinforcement learning process is configured to extract features from the weighted feedback, and generate at least one modification for the first image model based on the extracted features of the weighted feedback regarding the first reconstructed image;
generate a second image model by applying the at least one modification generated by the deep-reinforcement learning process to the first image model;
generate a second reconstructed image by applying the second image model to the entirety of the first patient scan data, the second reconstructed image comprising the imaging target at a second location that is different from the first location.
12. The system of claim 11 , wherein the feedback comprises one or more of system generated feedback or expert generated feedback.
13. The system of claim 12 , wherein the system generated feedback comprises one or more of an artifact reduction, an acquisition protocol, or a reconstruction protocol.
14. The system of claim 12 , wherein the expert generated feedback comprises one or more of a score indicative of an acceptability of the first reconstructed image or the at least one modification.
15. The system of claim 11 , wherein the feedback comprises a modified reconstructed image.
16. The system of claim 11 , wherein the imaging modality is configured to generate patient scan data during a second imaging period, and wherein the computer is further configured to:
receive the patient scan data generated during the second imaging period; and
generate a third reconstructed image by applying the second image model to the patient scan data generated during the second imaging period.
17. A non-transitory computer readable medium storing instructions configured to cause a computer system to execute the steps of:
receiving first patient scan data corresponding to a nuclear imaging scan of an individual;
receiving preference data associated with the first patient scan data;
selecting, by the system, a first image model from a plurality of image models based on at least a first portion of the preference data, and modifying, by the system, the first image model based on at least a second portion of the preference data;
generating a first reconstructed image by applying the first image model to an entirety of the first patient scan data, the first reconstructed image comprising an imaging target at a first location;
receiving a feedback regarding the first reconstructed image;
generating weighted feedback based on applying a first weight to a first portion of the feedback data and a second weight to a second portion of the feedback data;
providing the weighted feedback as an input to a deep-reinforcement learning process, wherein the deep-reinforcement learning process is configured to extract features from the weighted feedback, and generate at least one modification for the first image model based on the extracted features of the weighted feedback regarding the first reconstructed image;
generating a second image model by applying the at least one modification generated by the deep-reinforcement learning process to the first image model; and
generating a second reconstructed image by applying the second image model to the entirety of the first patient scan data, the second reconstructed image comprising the imaging target at a second location that is different from the first location.
18. The non-transitory computer readable medium of claim 17 , wherein the instructions are further configured to cause the computing system to execute the steps of:
receiving second patient scan data corresponding to one or more additional nuclear imaging scans; and
generating a third reconstructed image by applying the second image model to the second patient scan data.