Maziar Irannejad ¹ Iraj Abedi ² Mobina Naeemi sadigh ³ Vida Darbaghi Lonbani ³ Maryam Hassanvand ^3*

• ¹ Faculty of Electrical Engineering, Islamic Azad University of Najafabad, Najafabad, Isfahan, Iran
• ² Isfahan University of Medical Sciences, Isfahan, Isfahan, Iran
• ³ Department of Physics, Isfahan University of Technology, Isfahan, Isfahan, Iran

‌Background: The treatment planning process for brain tumors is a highly complex procedure due to the presence of many sensitive structures close to the target of radiation. Although the Monte Carlo algorithm used in treatment planning is very accurate, it is time-consuming. To solve this problem, deep learning models can be used to predict the dose distribution of brain tumors treated by intensity-modulated radiotherapy (IMRT). Purpose: In this article, the attention U-net model has been used to achieve higher accuracy in predicting the 3D dose of brain tumors. Methods: We investigate incorporating attention gates structure with the U-net model's architecture to enhance the prediction accuracy. The attention gate captures the significance of various spatial locations and can improve the performance of basic U-net models. Here, the data of 99 patients with glioma tumors who underwent manual planning and were treated with sixth-field IMRT using 6MV photon energy were utilized. The treatment plans used the Collapsed Cone Convolution algorithm to administer 60Gy in 30 fractions. The dose distribution images of 90 patients in different slices are used to train and validate the proposed model, and the images of 9 remaining patients are used for testing. Results: The model performance was evaluated by computing SSIM, Dice, IoU, and Bf1-score of the doses predicted images against clinical dose. This obtained accuracy based on the mentioned metrics are 0.829, 0.844, 0.704, and 0.63, respectively. The proposed model shows an improvement rate of about 0.01, 0.006, - 0.014, and 0.044 compared to the U-net model based on the mentioned criteria. Furthermore, the average absolute errors for the calculated DVH parameters of the proposed model are superior to those of the U-net model, with a range of difference of approximately (-0.218 to 2.55) percent.