Differentiable multi-agent actor-critic for multi-step radiology report summarization system
Systems and methods for using a differentiable multi-agent Actor-Critic (DiMAC) for multi-step radiology report summarization. The tasks of extracting salient sentences and phrases are divided across two collaborating agents that are trained end-to-end using reinforcement learning (RL).
1. A system for multi-step radiology report summarization, the system comprising:
a word extraction network configured to extract one or more words from a radiology report;
a sentence extraction network configured to extract one or more salient sentences from the radiology report based at least in part on the extracted one or more words; and
an abstractor network configured to condense one or more of the one or more extracted salient sentences into a concise summary of the radiology report;
wherein the word extraction network, the sentence extraction network, and the abstractor network are trained end to end using at least a critic and a communication channel, wherein the critic is configured to estimate a value function that is used to compute a policy gradient for the word extraction network and the sentence extraction network, and the communication channel is configured to pass gradient information between the word extraction network and the sentence extraction network.
2. The system of claim 1 , wherein the word extraction network and the sentence extraction network comprise a bi-directional LSTM based word encoder and a bi-directional LSTM sentence encoder respectively; wherein the bi-directional LSTM based word encoder is configured to obtain word representations from a FINDINGS portion of the radiology report and the bi-directional LSTM sentence encoder is configured to obtain sentence representations.
3. The system of claim 1 , wherein the abstractor network comprises a pointer generator network.
4. The system of claim 1 , wherein during training, the word extraction network and sentence extraction network are configured using multiple agent reinforcement learning.
5. The system of claim 1 , wherein during training, the word extraction network and sentence extraction network actions at a previous step are input into the communication channel that generates a sigmoidal, mt, that is input into the sentence extraction network, wherein 1−mt is input into the word extraction network.
6. The system of claim 1 , wherein during training, when the word extraction network or the sentence extraction network selects an action, a reward is provided to the word extraction network or the sentence extraction network to it based on an individual reward function computed specifically for either the word extraction network or the sentence extraction network.
7. The system of claim 6 , wherein the reward for the sentence extraction network is computed using ROUGE L recall by comparing extracted salient sentences and condensed salient sentences with ground truth data.
8. The system of claim 6 , wherein the reward for the word extraction network is computed by checking if an extracted word is in a given set of keywords.
9. The system of claim 1 , further comprising an output interface configured to provide the concise summary to a user.
10. A method for configuring a system for multi-step radiology report summarization, wherein the system comprises at least a sentence extraction network, a word extraction network, a critic network, and a communication channel, the method comprising:
inputting training data from a FINDINGS section of a radiology report to the sentence extraction network and the word extraction network;
computing states for the sentence extraction network, the word extraction network, and the critic network;
sampling actions by the sentence extraction network and the word extraction network;
computing rewards based on the actions and the states; and
updating the sentence extraction network, the word extraction network, and the critic network based on the computed rewards.
11. The method of claim 10 , wherein inputting, computing, sampling, computing, and updating is iteratively performed for a plurality of iterations.
12. The method of claim 11 , wherein when sampling actions by the sentence extraction network and the word extraction network, only one of the sentence extraction network or the word extraction network is active while the other is paused.
13. The method of claim 10 , wherein each of the sentence extraction network and the word extraction network selects one of its actions and communicates with the other respective network at every step.
14. The method of claim 10 , wherein the communication channel is configured to generate a sigmoidal, mt, that is input into the sentence extraction network wherein 1−mt is input into the word extraction network.
15. The method of claim 10 , wherein the word extraction network is configured to extract one or more words from the FINDINGS section based on a list of keywords.
16. The method of claim 15 , wherein the sentence extraction network is configured to extract one or more sentences from the FINDINGS section based in part on the extracted one or more words.
17. A method for automatic summarization of a FINDINGS section of a radiology report, the method comprising:
acquiring a radiology report that describes results of a medical procedure;
inputting the FINDINGS section of the radiology report into an automatic summarization system comprising at least a word extraction network, a sentence extraction network, a communications channel, and an abstractor, wherein the automatic summarization system is trained using differential Multi-agent Actor-Critic reinforcement learning;
outputting, by the automatic summarization system, an IMPRESSIONS section for the radiology report; and
providing the IMPRESSIONS section to a user.
18. The method of claim 17 , wherein the word extraction network is configured to extract one or more words from the FINDINGS section based on a list of keywords.
19. The method of claim 18 , wherein the sentence extraction network is configured to extract one or more sentences from the FINDINGS section based in part on the extracted one or more words.
20. The method of claim 17 , wherein the communications channel is configured to pass information between the word extraction network and the sentence extraction network.