In FDK reconstruction, distribution of noise power is different along the axial (i.e., high pass noise) and coronal slice (i.e., low pass or white noise), which may results in different detectability of same objects. In this work, we examined denoising performance of convolutional neural network trained using axial and coronal slice images separately, and how the direction of image slice affects the detectability of small objects in denoised images. We used the modified version of U-Net. For network training, we used Adam optimizer with a learning rate of 0.001, batch size of 4, and VGG loss was used. The XCAT simulator was used to generate the training, validation, and test dataset. Projection data was acquired by Siddon’s method for the XCAT phantoms, and different levels of Poisson noise was added to the projection data to generate quarter dose and normal dose CT images, which were then reconstructed by FDK algorithm. The reconstructed quarter dose and normal dose CT images were used as training, validation, and test dataset for our network. The performance of denoised output images from U-Net-Axial (i.e., network trained using axial images) and U-Net-Coronal (i.e., network trained using coronal images) were evaluated using structural similarity (SSIM) and mean square error (MSE). The results showed that output images from both U-Net-Axial and U-Net-Coronal shows the improved image quality compared to quarter dose images. However, it was observed that the detectability of small objects were higher in U-Net-Coronal.