Exploring the impact of field shape on predicted dose distribution in breast cancer patients using deep learning in radiation therapy

Geometrical information such as field shape is essential for dose calculation in radiation therapy. However, new methods of dose prediction based on deep learning only use CT images and contouring of patients. The aim of this study is to develop a deep learning method for three-dimensional dose prediction for breast cancer radiotherapy using field shape as well as CT and contouring images of patients. A total of 150 breast cancer patients' data treated with 3D conformal radiation therapy (CRT) technique were used to train and test 3D-U-Res-F-Net model. The model inputs were CT images, patients' contouring and field shape. The network output was the corresponding the patients’ dose distribution. Then, the dosimetric parameters extracted from the dose volume histograms related to the planned and predicted dose distributions were compared and were statistically analysed with the paired-samples t-test. These parameters include Dmean and Dmax for planning target volume and organs at risk, D95%, D50%, V47.5Gy for PTV, V30Gy and V25Gy for heart and V20Gy for left lung. The gamma index with 3%/3 mm criteria was calculated for PTV and OARs. The average absolute difference of the Dmean relative to the prescribed dose for the PTV, heart, left lung, right lung and spinal cord were 1.37 ± 0.4%, 2.02 ± 2%, 2.12 ± 1%, 0.37 ± 0.3% and 0.41 ± 0.3%, respectively. The average absolute difference of the Dmax relative to the prescribed dose for the PTV and OARs varied from 0.23 ± 0.2% to 2.04 ± 2%. The 3D gamma pass rate with 3mm/3% criteria for planning target volume, heart, left lung, right lung, spinal cord and body were 89 ± 3%, 91 ± 5%, 93 ± 5%, 99 ± 3%, 99 ± 1% and 96 ± 2% respectively. The results of the proposed model demonstrate that there is no significant difference between the predicted and planned dose distributions. The deep learning model can directly predict the 3D dose distribution faster than treatment planning systems. Further studies with more patients and on other cancer sites are essential to fully validate the proposed method. •A 3D deep learning model was used to predict the dose distribution of 3D-CRT technique.•The performance of the 3D-U-Res-F-Net model show that the similarity between the predicted and planned dose is appropriate.•Adding the shape of the field in the training process of the DL model improves the accuracy of the model in dose prediction.