We investigated the feasibility of the generation of synthetic CT (sCT) from CBCT images with deep learning and the dose evaluation for CBCT-guided breast cancer adaptive radiotherapy. A total of sixty-eight patients receiving radiotherapy after breast-conserving surgery were retrospectively included in this study. We compared the performance of three deep-learning methods in generating sCT from CBCT, including U-Net, Cycle generative adversarial network (CycleGAN) and pix2pix. The original treatment plan was transferred to sCT keeping the same parameters. The dosimetric evaluation was performed by a quick dose recalculation on sCT based on gamma analysis. The U-Net model obtained the lowest mean absolute error (MAE) within the body, clinical target volume (CTV) and organs at risk (OARs), with 62.53 +/- 9.14 HU within the body, 35.99 +/- 6.32 HU within tumor bed, and 30.15 +/- 6.36 HU within CTV. In terms of dose comparison, the gamma pass rates under 3%/3 mm and 2%/2 mm criteria were 91.40 +/- 3.52% and 85.95 +/- 4.75% for the U-Net model, whereas 89.50 +/- 3.46% and 83.65 +/- 4.00% for the pix2pix model, and 89.84 +/- 3.47% and 83.69 +/- 4.28% for the CycleGAN model, respectively. The sCT images generated by the U-Net model can provide higher image similarity and dosimetric accuracy than those generated by the pix2pix and CycleGAN models. The approach could be used to realize accurate dose calculation for breast cancer adaptive radiotherapy based on CBCT.