K. Zhang1, J. Shen1, B. Yang2, Y. Liang2, X. Meng1, X. Hou3, K. Hu2, J. Qiu2, and F. Q. Zhang4; 1Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China, 2Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China, 3Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of medical Sciences & Peking Union Medical College, Beijing, China, 4Department of Radiation Oncology, Peking Union Medical College Hospital. Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China
Purpose/Objective(s): The study developed and validated a deep learning-based dose prediction model for left-sided breast cancer patients undergoing whole-breast radiotherapy with deep inspiration breath-hold (DIBH). The aim was to select patients who would benefit significantly from DIBH, to avoid its unnecessary use of DIBH, reduce clinical costs, and enhance the efficiency of clinical treatment. Materials/
Methods: The study prospectively involved patients with left-sided breast cancer who had undergone breast-conserving surgery and were scheduled for radiotherapy. CT images were acquired under both the DIBH and free-breathing (FB). All patients received dynamic intensity-modulated radiotherapy with a prescribed dose of 42.4 Gy in 16 fractions, which included 4 tangential beams. The patients were randomly divided into training and validation sets in a 4:1 ratio. The study utilized a 3D U-Net-based deep learning model. The model takes 3D image information as input and outputs a 3D dose field. The primary data included CT simulation images, regions of interest (ROIs) sketched manually and automatically segmented, and the dosage information of treatment plans manually prepared by experts. To predict the irradiation dose to the heart under FB and select patients using 200 cGy as a threshold, we developed model A, which used the automatically segmented contours of the left breast, heart, and body as inputs and was trained with data under both FB and DIBH and tested using data under FB. To predict the heart dose reduction under DIBH compared with FB, we developed model B which utilized automatically segmented contours under FB as inputs. The training target was the heart dose difference between FB and DIBH. Results: The study involved 51 left-sided breast cancer patients. The training set consisted of 80 manually prepared treatment plans for 40 patients, under both DIBH and FB. The test set comprised 22 plans for 11 patients. Model A predicted the mean cardiac dose under FB with an average error of 52.82 cGy. To determine whether DIBH is necessary, a mean heart dose of 200 cGy under FB serves as a threshold. If the mean cardiac dose is below 200 cGy, DIBH is not necessary as the reduction in the mean heart dose is small. The test set showed that Model A achieved 100% classification accuracy with 200 cGy as the threshold. The predicted dose difference between DIBH and FB by Model B had a mean absolute error of 59.53 cGy. Model B accurately predicts the reduction of the mean heart dose under DIBH and provides a basis for physicians to select patients suitable for DIBH treatment. Conclusion: The study applied automated segmentation techniques and three-dimensional dose prediction methods to rapidly and accurately screen breast cancer patients suitable for DIBH. Our model predicted the cardiac irradiated dose under FB and assessed the reduction of cardiac dose under DIBH using the automatic segmented image under FB. The model has significant implications for clinical decision-making and the personalized design of DIBH therapy.
