You have done your CG reconstruction assuming symmetry about the center of the matrix rather than the center of k-space.  For your example the center of k-space is at (129,129). If you use the correct point of symmetry your CG reconstructed image will be almost identical to your reference image.  Because you used the wrong point of symmetry your CG reconstructed k-space has a discontinuity in it and that is causing the ghosting.

For the images from your sample code the zero-filled image looked better than the ghosted CG image because you replaced part of the high spatial frequency information in k-space with zeros, resulting in a smoothed image.

In this example, k-space is also 256*256. Where is the center locate?

Can any body help to fix the problem and share us the correct code? 

When you do an fft the center of k-space is at (1,1).  After you do the ifftshift for each dimension, the center is at (129,129).  In your code you assume that the center of k-space is at (128.5,128.5).  For example you assume that there is conjugate symmetry between column 128 and 129.  In reality though column 128 has conjugate symmetry with column 130 rather than 129.  Column 256 has conjugate symmetry with column 2 rather than column 1 as you have assumed.  You replaced columns 1-112 with columns  145-256, but column 112 has conjugate symmetry with column 146 rather than 145.  You should modify the following line

tmp = kspRef(:,end-floor((1-pf)*N)+1:end);

with

tmp = kspRef(:,end-floor((1-pf)*N)+2:end);

and place these columns into columns 2 thru 112 after taking the conjugate and flipping.

In addition when you did the following

C = flipud(B);

you assume symmetry between row 128 and 129 etc.  Instead you need to do

C = circshift(flipud(B),1);

With this change row 129 in B is still row 129 in C.

If you make these changes you should get the attached picture as output.

The result is right now! And the explanation is perfect! THX!