<p>Low-dose computed tomography (LDCT) is essential for reducing patient radiation exposure but often suffers from high noise and streak artifacts. The vast majority of publications on LDCT denoising focus on image-domain approaches. This is mainly because noise reduction in the projection domain is considerably more challenging, and access to raw projection data from real CT systems is often limited. In this study, we compared the performance of various deep learning–based denoising models applied to 2D cone-beam CT projections. Four different models were investigated, including a U-Net, an edge-enhancement dense feature propagation network (EDCNN), a generative adversarial network (GAN), and a diffusion-based model. The comparison was conducted focusing on two different training strategies. For evaluation, both real 2D cone-beam projection data and projections of mathematical phantoms were used. Training was performed exclusively on projections of realistic mathematical phantoms. These phantoms were forward-projected using the cone-beam geometry of a real CT system, and a validated noise model was applied to simulate Poisson and Gaussian noise components. The comparative analysis was carried out on images reconstructed using filtered back projection (FBP) on PI-lines. This work highlights the importance of physical modeling, data consistency, and the training strategy of deep learning–based algorithms for robust, generalizable LDCT denoising.</p>