Open sailuluo opened 3 years ago
I train with the default set, the DC_CNN result far from paper's result
So, u mean that the real part has higher PSNR than the result of network? Would u like to share some more like modified code?
For you second question, do u mean that DC_CNN should result to higher PSNR? But the default config file in this repo works for the proposed model not DCCNN, or u just use the config file from storage repo? Anyway, the PSNR result is too higher than exception, no mattar DCCNN or proposed. So would u like to share more information, like the config file, experiments? Thank u so much.
hello, I finally find the reason. the dataset processed by the convert code not same as that you provide. The result is normal after using the dataset you provided
for cardiac dataset. and use default.nii. when validation, I modify part of the code to evaluate the zero-fillied image with ground truth. result is: Evaluating 0300 psnr(before|after) = 36.63|35.94 ssim = 0.9410|0.9089 Done
compare with after, before just missing an imaginary part.
why the zero-fillied image without reconstruction can achieve higher PSNR score? then I evaluate 10% sample ratio, 5% sample ratio, all PSNR and SSIM are higher.