Deep Learning-Based Automatic Delineation of Target Volumes and Organs at Risk of Breast Cancer for On-Line Dosimetric Evaluation.

In this study, we developed a deep learning model to achieve automatic multi-target delineation on planning CT (pCT) and daily Cone-Beam CT (CBCT). To improve the image quality of the CBCT for accurate target delineation and dose calculation, we introduce an unsupervised learning model to generate the artifact-free synthetic CT (sCT) from the CBCT. The geometric and dosimetric impact of the model on planning CT (pCT) and synthetic CT (sCT) was evaluated for breast cancer adaptive radiation therapy.We retrospectively analyzed 1127 patients treated with radiotherapy after breast-conserving surgery from two medical institutions. The CBCT for patient setup acquired utilizing breath-hold guided by optical surface monitoring system was used to generate sCT with the generative adversarial network. Organs at risk (OARs) and target volumes including tumor bed (TB) and clinical target volume (CTV) for breast cancer were delineated automatically with 3D U-Net model on pCT and sCT. The automatic delineation was compared with manually delineated contours to evaluate the performance with geometric metrics, including Dice similarity coefficient (DSC) and 95% Hausdorff Distance (HD95). The treatment plan which was transferred to the sCT from the pCT was generated with the same planning parameters as the original pCT-based plan. The dosimetric evaluation was performed by a quick dose recalculation on sCT relying on gamma analysis and the dose-volume histogram (DVH) parameters. The automatically delineating CTV on sCT which was rigidly registered to pCT was compared with manually delineating CTV on pCT to obtain DSC-CTV. The relationship between the ∆D95, ∆V95 and DSC-CTV was assessed to quantify the clinical impact of the geometric changes of CTV.The range of the DSC and HD95 were 0.73-0.97, 2.22-9.36mm for pCT, 0.63-0.95, 2.30-19.57mm for sCT from institution A, 0.70-0.97, 2.10-11.43mm for pCT from institution B respectively. The quality of sCT was excellent with an average mean absolute error (MAE) of 71.58 ± 8.78HU. The mean gamma pass rate (3%/3 mm criterion) by comparing the dose on sCT with that on original pCT was 91.46 ± 4.63%. DSC-CTV down to 0.65 accounted for a variation of more than 6% of V95 and 3Gy of D95. DSC-CTV up to 0.80 accounted for a variation of less than 4% of V95 and 2Gy of D95. The mean ∆V95 of CTV was less than 6%. The mean ∆V95 of TB was more than 8%. The mean ∆D90/∆D95 of CTV and TB were less than 2Gy/4Gy, 4Gy/5Gy for all the patients. The cardiac dose difference in left breast cancer was bigger than that in right breast cancer.This study demonstrates that highly accurate multi-target delineation and dose calculation are achievable using the daily CBCT image via deep learning. The results show that dose distribution needs to be considered to evaluate the clinical impact of geometric variations to decide whether to re-plan during breast cancer radiotherapy.

Duke Scholars

Author Jing Cai Radiation Oncology