Giters

chunmeifeng / MINet

【MICCAI 2021, Early accept】Multi-Contrast MRI Super-Resolution via a Multi-Stage Integration Network

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

bicubic interpolation results

lilygeorgescu opened this issue · comments

commented

Hi,

I computed the PSNR and SSIM values using your code (common/evaluate.py) on the fastMRI dataset (on the test split provided by you) using the singlecoil images and I obtained the following results:

Bicubic interpolation method on PDWI (x2): PSNR = 37.1 +/- 4.023 SSIM = 0.8996 +/- 0.07312
Bicubic interpolation method on FS-PDWI (x2): PSNR = 31.86 +/- 4.345 SSIM = 0.728 +/- 0.1362

These numbers are very different from those reported in your paper. Did you use any other pre-processing step before applying bicubic interpolation? Or did you use other evaluation protocol?

Thank you,
Best regards,
Lili Georgescu

commented
commented

Hi,

First, I want to thank you for taking the time to answer me. I really appreciate it.

I used the code from your repo (common/evaluate.py), and I replaced the loading of the reconstruction images:

so instead of

recons = recons['reconstruction'][()]

I did this:

recons = resize(target.copy())

def resize(images):
    new_images = []
    for image in images:
        h, w = image.shape

        res_image = cv.resize(
            cv.resize(image, None, fx=0.5, fy=0.5, interpolation=cv.INTER_CUBIC),
            (w, h),  interpolation=cv.INTER_CUBIC)

        new_images.append(res_image)

    return np.array(new_images)

The resize function downsamples the image using bicubic interpolation, then it upsamples the image back to its original same (also using bicubic).

Did you do anything different when computing the results for bicubic interpolation?

I used the images from the singlecoil folder downloaded from the original website.

Thank you,
Best regards,
Lili Georgescu

commented