In PET/CT imaging, CT is used for PET attenuation correction (AC). Mismatch between CT and PET due to patient body motion results in AC artifacts. In addition, artifact caused by metal, beam-hardening and count-starving in CT itself also introduces inaccurate AC for PET. Maximum likelihood reconstruction of activity and attenuation (MLAA) was proposed to solve those issues by simultaneously reconstructing tracer activity ($\lambda$-MLAA) and attenuation map ($\mu$-MLAA) based on the PET raw data only. However, $\mu$-MLAA suffers from high noise and $\lambda$-MLAA suffers from large bias as compared to the reconstruction using the CT-based attenuation map ($\mu$-CT). Recently, a convolutional neural network (CNN) was applied to predict the CT attenuation map ($\mu$-CNN) from $\lambda$-MLAA and $\mu$-MLAA, in which an image-domain loss (IM-loss) function between the $\mu$-CNN and the ground truth $\mu$-CT was used. However, IM-loss does not directly measure the AC errors according to the PET attenuation physics, where the line-integral projection of the attenuation map ($\mu$) along the path of the two annihilation events, instead of the $\mu$ itself, is used for AC. Therefore, a network trained with the IM-loss may yield suboptimal performance in the $\mu$ generation. Here, we propose a novel line-integral projection loss (LIP-loss) function that incorporates the PET attenuation physics for $\mu$ generation. Eighty training and twenty testing datasets of whole-body 18F-FDG PET and paired ground truth $\mu$-CT were used. Quantitative evaluations showed that the model trained with the additional LIP-loss was able to significantly outperform the model trained solely based on the IM-loss function.