Room: AAPM ePoster Library
Purpose: set up an intelligent treatment planning program for left breast cancer.
Methods: The treatment plans belonging to 200 patients who suffered left breast cancer were adopted. (2) We set the threshold value by referring to RTOG-1005 and RTOG-1304. (3) The deep convolution Generate Adversarial Network (GAN) has been used as the engine for AI dose prediction. The Generator(G)and Discriminator(D)is working via following a standard formula.(4) U-net neural network model was adopted for DVH prediction. The input of U-net network is 256*256*3 CT contour map, and the output is the dose distribution map of corresponding single-pass. (5) For automatic optimization algorithm, a series of linear objective functions were used as a plan optimization engine to constitute an inverse plan optimization problem. During calculations, we turn the optimization problem into a form that optimizes the absolute dual gap.
Results: left chest wall with supraclavicular case of breast cancer: from the view of DVH generated by the optimization engine, the results on PTV and OAR from automatic plans were better than those from manual plans. From the predicted dose distribution graph, our results from AI dose prediction system were most equivalent to or better than those from manual plans.
Conclusion: left breast chest wall with supraclavicular fields of left breast cancer, the DVH generated by our automatic plan optimization engine were better than those from the manual plan. There is no bright spot or noise on the fluence map, which has good smoothness and can guarantee the execution of the plan.
Breast, Dose Volume Histograms