A CNN-based dose prediction method for brachytherapy treatment planning of patients with cervical cancer

Brachytherapy (BT) plays a crucial role in cervical cancer treatment. This study aimed to develop a 3D dose prediction model for cervical BT using Convolutional Neural Network (CNN). In this study, we introduced a dose prediction model guided to generate dose distributions with explicit anatomical mask guidance. The model encompassed 224 clinical cases, including 190 for training-validation and 34 for testing. For performance evaluation, DVH metrics and 3D Gamma analysis were employed. The results were compared with those obtained using a 3D U-net model. DVH metrics for the test set, including HRCTV D90, HRCTV D95, HRCTV D100, bladder D2CC, sigmoid D2CC, rectum D2CC, and intestine D2CC, yielded values of 5.44 ± 0.91, 5.05 ± 0.88, 3.34 ± 0.79, 4.39 ± 1.53, 3.24 ± 1.31, 3.03 ± 1.87, and 2.71 ± 1.79, respectively. The DVH metrics of dose differences between the predicted dose distribution and the ground-truth plan were 0.63 ± 0.63, 0.60 ± 0.61, 0.53 ± 0.61, 1.21 ± 0.85, 0.71 ± 0.61, 1.16 ± 1.09, and 0.86 ± 0.58, respectively. The 3D gamma passing rates for the 3%/3 mm criteria of HRCTV, bladder, sigmoid, rectum, and intestine were 0.95 ± 0.04, 0.99 ± 0.02, 1.00 ± 0.02, 1.00 ± 0.01, and 1.00 ± 0.00, respectively. The 3D BT dose prediction system, based on a 3D anatomical mask-guided deep learning network, could accurately generate 3D dose distributions, offering decision support for automatic clinical BT treatment planning.