search

fabianhielscher / AUTOMAP

This is an implementation of AUTOMAP which reconstructs MRI-images with undersampled k-space data
7 stars 3 forks source link

"log" and "lin" #1

Open JiannpingXu opened 3 years ago

JiannpingXu commented 3 years ago

Hi, may I ask what "log" and "lin" mean in your code? Thank you very much!

fabianhielscher commented 3 years ago

hi,

during preprocessing the discrete 2d fourier transform is applied to 2d image data to simulate k-space data, such as: img_f = np.fft.fft2(pixel) # FFT

After normalization, these values are not evenly distributed. I tought i might get better performance results if the input data is consistantly scaled logarithmically before applying normalization, such as: img_f_log = img_f / np.power(abs(img_f)+1, 1/np.exp(1))

To distinguish logarithmically scaled data from non-scaled data, I called them log (logarithmically scaled) and lin (linear, non-scaled). It turned out, that the performance decreased with the scaling idea, so I used lin data for training.

I hope that helped.