We use small subset of fastMRI singlecoil knee dataset. Dataset consist only from PD, 3T scans and slices selected only at center of knee (dataset without slices on knee borders)
Link to dataset (4 Gb) You should have h5py > 3.2 and gdown > 3.12, you can update like that:
python -m pip install gdown==3.12.2
python -m pip install h5py==3.2.1
Gaussian
from k_space_reconstruction.datasets.fastmri import FastMRITransform, RandomMaskFunc
transform = FastMRITransform(
RandomMaskFunc([0.08], [4]),
noise_level=100,
noise_type='normal'
)
Salt
from k_space_reconstruction.datasets.fastmri import FastMRITransform, RandomMaskFunc
transform = FastMRITransform(
RandomMaskFunc([0.08], [4]),
noise_level=5e4,
noise_type='salt'
)
Gaussian + Salt
from k_space_reconstruction.datasets.fastmri import FastMRITransform, RandomMaskFunc
transform = FastMRITransform(
RandomMaskFunc([0.08], [4]),
noise_type='normal_and_salt'
)
Unet16
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8053 | 0.0099 | 31.8321 |
gaussian | 0.7210 | 0.0142 | 30.3041 |
salt&pepper | 0.6806 | 0.0207 | 28.9547 |
gaussian + salt&pepper | 0.6807 | 0.0189 | 28.2086 |
Attention-Unet16
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.4939 | 0.2021 | 18.6097 |
gaussian | 0.4426 | 0.2127 | 18.3985 |
salt&pepper | 0.3875 | 0.3336 | 16.6910 |
gaussian + salt&pepper | 0.3624 | 0.3872 | 16.1630 |
Cascade-5x-Unet16-noDC
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8013 | 0.0097 | 31.9888 |
gaussian | 0.7058 | 0.0151 | 30.0426 |
salt&pepper | 0.6169 | 0.0394 | 26.5311 |
gaussian + salt&pepper | 0.6086 | 0.0574 | 25.8732 |
Cascade-5x-Unet16-DCL
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8444 | 0.0069 | 33.4667 |
gaussian | 0.7388 | 0.0117 | 31.2349 |
salt&pepper | 0.6156 | 0.0262 | 28.2839 |
gaussian + salt&pepper | 0.5892 | 0.03026 | 27.8198 |
gaussian_400 | 0.6035 | 0.0242 | 28.3150 |
gaussian_400 + salt&pepper | 0.5262 | 0.0419 | 26.2551 |
Cascade-5x-Unet16-DC
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8508 | 0.0064 | 33.8926 |
gaussian | 0.6747 | 0.0171 | 29.8601 |
salt&pepper | 0.3862 | 0.1945 | 21.1475 |
gaussian + salt&pepper | 0.2580 | 0.3408 | 18.0605 |
Cascade-5x-Unet16-DC-AF
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8419 | 0.0071 | 33.3439 |
gaussian | 0.7254 | 0.0125 | 30.9641 |
salt&pepper | 0.5208 | 0.0584 | 25.8676 |
gaussian + salt&pepper | 0.5130 | 0.0517 | 25.8669 |
Cascade-5x-Unet16-DC-Super-AF
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8428 | 0.0070 | 33.4662 |
gaussian | 0.7508 | 0.0111 | 31.4516 |
salt&pepper | 0.8347 | 0.0082 | 32.8732 |
gaussian + salt&pepper | 0.7359 | 0.0133 | 30.7945 |
Cascade-5x-Unet16-TDC-FtF-Super-AF
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8509 | 0.0063 | 33.9353 |
gaussian | 0.7510 | 0.0109 | 31.5610 |
salt&pepper | 0.8437 | 0.0071 | 33.4054 |
gaussian + salt&pepper | 0.7361 | 0.0121 | 31.1309 |
DnCNN
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.7742 | 0.0148 | 30.1280 |
gaussian | 0.6676 | 0.0215 | 28.5129 |
salt&pepper | 0.3955 | 0.0827 | 23.2475 |
gaussian + salt&pepper | 0.3467 | 0.1060 | 22.3551 |
Cascade-5x-DnCNN-DC
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8394 | 0.0072 | 33.2991 |
gaussian | 0.6639 | 0.0182 | 29.5485 |
salt&pepper | 0.3112 | 0.2565 | 19.3469 |
gaussian + salt&pepper | 0.2022 | 0.5462 | 16.1158 |
Cascade-5x-DnCNN-DCL
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.8325 | 0.0078 | 32.9421 |
gaussian | 0.7098 | 0.0138 | 30.4711 |
salt&pepper | 0.4911 | 0.0420 | 25.9640 |
gaussian + salt&pepper | 0.4276 | 0.0693 | 24.4179 |
Cascade-5x-DnCNN-NoDC
noise | SSIM | NMSE | PSNR |
---|---|---|---|
none | 0.7712 | 0.0154 | 29.9269 |
gaussian | 0.6601 | 0.0227 | 28.3064 |
salt&pepper | 0.4276 | 0.0654 | 23.9284 |
gaussian + salt&pepper | 0.4551 | 0.0717 | 23.9338 |
* - need revision