Enhancement of low-quality noisy CBCT images plays a crucial role in the image-guided radiotherapy (IGRT) treatment of breast cancer surgery patients. The objective of this paper is to identify corresponding pairs of pCT and CBCT images, prepare a supervised learning dataset and evaluate the performance of four state-of-the-art deep CNN models in generating high-quality improved CBCT (iCBCT) images. In this paper, pCT and CBCT image samples are acquired from 5 Breast-conserving surgery (BCS) patients of a leading cancer institute in India. The histogram comparison technique matches corresponding pCT and CBCT image pairs to prepare labelled datasets. Models such as Deep CNN Autoencoder, U-Net, CBDNet, and RIDNet are trained and tested with these paired normalized pCT and CBCT images to predict iCBCT images. The quality of these iCBCT images is then compared with that of the pCT images using performance measures such as Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Feature Similarity Index Measure (FSIM). The average percentage improvement of FSIM between corresponding pCT and CBCT versus pCT and iCBCT for Autoencoder, U-Net, CBDNet and RIDNet models is 55.2%, 28.28%, 55.1%, and 17.24%, respectively. Both Autoencoder and CBDNet achieve a 21.62% higher FSIM value than the U-Net model and U-Net achieve an 8.8% higher FSIM value than the RIDNet model. Autoencoder and U-Net models achieve 7.5% and 4.5% higher SSIM values than CBDNet and RIDNet models, respectively.