Breast cancer dose distribution prediction based on deep joint learning.

Background And Purpose: In radiotherapy planning, achieving a balance between the dose distribution for the planning target volume (PTV) and organs-at-risk (OARs) is critical. This study introduces a joint learning mechanism to progressively refine predictions for PTVs and OARs, aiming to improve the accuracy of dose prediction for breast cancer.

Materials And Methods: The proposed model begins with constructing a dose prediction network tailored for the PTVs. The output from this network serves as an additional input for a subsequent network, which predicts the overall dose distribution map. The framework is further enhanced by integrating distance transformation and a dual attention module into the network to improve prediction accuracy. The experimental dataset comprises records from 307 postoperative breast cancer patients who underwent Intensity-Modulated Radiation Therapy (IMRT). The dataset includes contours of OARs, PTVs, and dose files.

Results: The proposed method demonstrates outstanding accuracy in dose distribution prediction. Compared to the baseline 3DU-Net method, our method improves the dose score by 0.169 Gy, the DVH score by 0.529, ΔPCI by 0.025, PSNR by 1.382 dB, SSIM by 0.017, ΔD95 by 1.779, ΔD98 by 1.153, ΔD99 by 0.547, and ΔDmax by 1.102. Moreover, compared to cascaded networks that predict PTV and OAR distribution maps separately, the proposed network, which directly predicts the entire dose distribution map in the second stage, demonstrates superior performance.

Conclusion: The proposed method can generate highly accurate three-dimensional dose distribution maps, which can be utilized for initialization to enhance the efficiency of radiotherapy planning.