MORE: Multi-Organ medical image REconstruction

Shaokai Wu
Yapan Guo*
Yanbiao Ji
Jing Tong
Yue Ding
Yuxiang Lu
Mei Li
Suizhi Huang
Hongtao Lu*

ACM MultiMedia 2025 Dataset

Overview

We propose a comprehensive dataset for the purpose of CT reconstruction. Compared to previous dataset, our dataset has the following advantages:
(1) Larger. There are 135 CT instances with 65,575 CT scans.
(2) Multi-Organ. Containing 9 types of different anatomy.
(3) Multi-Lesion. Containing 15 types of lesions.

Image 1

MORE Dataset

Our dataset is available at Huggingface.


Below is some samples of MORE dataset:

GIFT Model



Innovation Comparison

Neural-network-based DLR methods can be typically divided into two categories:

Direct Learning methods learns the mapping from raw measurements to reconstructed images directly by the neural-networks, such as AUTOMAP and iRadonMAP.







Indirect Learning methods treat CT reconstruction as a high-level denoising problem, processing raw sinogram data through FBP or IR to the image domain, and aim to train a network to map from low-dose to full-dose sinograms or images. Typical methods include RED-CNN, FBPConvNet, etc.







The most advanced DLR methods are based on diffusion models, such as DiffusionMBIR, SWORD, etc. These methods learn the score function of the posterior distribution of the image given the sinogram data, and can be used to reconstruct high-quality images from low-dose sinograms. The diffusion models have shown superior performance in CT reconstruction tasks. The following figure is sourced from SWORD







Compared to these methods, our GIFT model does not rely on the training of any neural networks and can directly reconstruct high-quality images under the supervision of low-dose sinograms.

On the other hand, 3D Gaussian-Based method R2-Gaussian is a recent method that uses a 3D Gaussian representation to model the CT image, and can achieve high-quality reconstruction results. R2-Gaussian employs rectified radiative Gaussian kernels to represent 3D density fields and culls the Gaussians in 3D tiles shaped 8x8x8 to reduce the computational cost. The following figure is sourced from the pipeline of R2-Gaussian:







Compared to R2-Gaussian, our GIFT model (1) represents all Gaussians in 3-sigma truncated format to reduce the computational complexity, (2) directly reconstructs the whole CT volume and optimizes all involved Gaussians. Extensive Benchmark demonstrates that our GIFT model is a stronger baseline that will benefits the community.
Below is our distinctive pipeline of GIFT from previous neural-network based methods:


MORE Benchmark

Emphysema Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 29.58 0.714 28.44 0.644 27.06 0.623 27.27 0.588
MCG 32.72 0.820 32.84 0.821 34.47 0.843 32.90 0.820
DiffusionMBIR 32.58 0.933 32.64 0.936 32.45 0.932 32.24 0.932
SWORD 35.38 0.879 34.52 0.864 33.78 0.849 32.30 0.827
FBP × 18.55 0.365 16.29 0.293 14.77 0.248 12.03 0.193
NeRP × 25.41 0.744 25.21 0.735 25.40 0.745 25.39 0.745
R2-Gaussian × 38.88 0.943 38.52 0.939 37.95 0.932 37.69 0.928
GIFT × 39.47 0.950 39.04 0.946 38.42 0.941 38.04 0.937
Ureteral Calculi Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 37.04 0.901 35.63 0.913 32.07 0.759 31.46 0.844
MCG 37.94 0.901 37.99 0.901 38.04 0.902 38.05 0.902
DiffusionMBIR 38.37 0.968 38.24 0.967 38.13 0.967 38.90 0.966
SWORD 42.35 0.973 40.93 0.967 39.42 0.960 37.63 0.947
FBP × 23.09 0.515 19.42 0.462 16.89 0.416 14.02 0.355
NeRP × 26.91 0.801 26.68 0.789 26.95 0.802 26.66 0.785
R2-Gaussian × 41.37 0.971 41.09 0.966 40.05 0.962 39.45 0.956
GIFT × 43.43 0.982 42.24 0.980 40.82 0.976 40.11 0.975
Rib Fracture Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 29.61 0.707 28.97 0.682 27.94 0.658 27.97 0.585
MCG 34.81 0.851 34.94 0.852 34.96 0.853 35.07 0.854
DiffusionMBIR 34.64 0.950 34.64 0.952 34.54 0,951 34.35 0.950
SWORD 36.51 0.877 35.90 0.864 35.53 0.855 34.76 0.838
FBP × 19.33 0.388 16.64 0.324 14.76 0.280 12.69 0.230
NeRP × 25.77 0.778 25.10 0.744 25.63 0.771 25.60 0.769
R2-Gaussian × 40.72 0.960 39.48 0.952 38.65 0.948 37.68 0.943
GIFT × 42.43 0.972 41.05 0.962 40.01 0.953 39.43 0.948
Appendicitis Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.96 0.904 35.54 0.906 31.30 0.838 32.59 0.854
MCG 38.76 0.908 38.96 0.909 38.97 0.897 38.36 0.899
DiffusionMBIR 38.34 0.960 38.28 0.959 38.24 0.966 38.00 0.967
SWORD 44.18 0.976 42.62 0.971 40.85 0.964 37.79 0.949
FBP × 23.37 0.516 19.63 0.462 18.17 0.427 14.67 0.366
NeRP × 27.15 0.821 27.25 0.817 27.38 0.819 27.28 0.817
R2-Gaussian × 41.47 0.964 40.78 0.959 40.16 0.954 39.79 0.948
GIFT × 42.03 0.981 41.63 0.981 41.15 0.979 40.22 0.976
Pneumonia Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 31.78 0.733 30.43 0.672 29.22 0.680 27.82 0.578
MCG 32.87 0.810 33.05 0.813 33.19 0.814 33.33 0.815
DiffusionMBIR 33.34 0.954 33.26 0.953 33.10 0.952 32.86 0.951
SWORD 39.69 0.901 38.75 0.887 38.02 0.875 36.40 0.850
FBP × 17.57 0.323 15.73 0.264 14.66 0.229 12.73 0.182
NeRP × 25.52 0.694 26.16 0.733 25.93 0.722 25.64 0.701
R2-Gaussian × 39.73 0.953 38.96 0.949 38.40 0.944 37.92 0.938
GIFT × 41.77 0.967 40.96 0.962 40.31 0.956 39.11 0.946
Cerebral Hemorrhage Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 35.47 0.895 33.29 0.864 30.46 0.786 29.26 0.766
MCG 39.14 0.898 39.23 0.899 39.32 0.899 39.31 0.899
DiffusionMBIR 39.04 0.969 39.29 0.973 39.05 0.971 38.53 0.969
SWORD 34.90 0.742 33.50 0.740 31.86 0.737 29.57 0.732
FBP × 24.13 0.526 21.54 0.490 19.70 0.460 17.52 0.413
NeRP × 25.38 0.789 25.98 0.804 25.02 0.760 24.23 0.764
R2-Gaussian × 40.97 0.968 40.54 0.964 39.88 0.960 38.79 0.955
GIFT × 43.71 0.984 42.94 0.981 41.68 0.978 40.56 0.974
Kidney Stones Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.65 0.882 34.70 0.909 31.98 0.802 30.89 0.798
MCG 38.16 0.909 38.43 0.911 38.49 0.912 38.67 0.913
DiffusionMBIR 28.84 0.964 38.92 0.966 38.79 0.964 38.54 0.964
SWORD 43.58 0.980 42.27 0.976 40.95 0.971 39.51 0.961
FBP × 22.88 0.483 19.39 0.439 16.27 0.398 13.52 0.341
NeRP × 26.17 0.767 26.25 0.773 26.11 0.772 26.16 0.776
R2-Gaussian × 44.00 0.983 42.94 0.978 41.97 0.975 40.26 0.969
GIFT × 44.37 0.988 43.45 0.987 42.99 0.986 41.20 0.982
Fatty Liver Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.60 0.857 35.64 0.876 32.48 0.743 32.73 0.836
MCG 37.97 0.897 38.07 0.897 38.12 0.898 38.14 0.898
DiffusionMBIR 38.04 0.961 37.95 0.960 37.86 0.960 37.68 0.959
SWORD 43.47 0.973 42.21 0.968 40.76 0.961 38.45 0.948
FBP × 22.29 0.482 18.10 0.431 16.54 0.395 13.87 0.342
NeRP × 26.89 0.785 27.27 0.808 26.81 0.784 26.93 0.792
R2-Gaussian × 42.58 0.977 42.03 0.974 41.36 0.969 40.80 0.965
GIFT × 44.46 0.987 43.96 0.986 43.47 0.985 42.54 0.983
Gallbladder Stones Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.15 0.892 35.59 0.913 32.41 0.797 31.80 0.868
MCG 38.13 0.897 38.47 0.901 38.01 0.897 37.95 0.899
DiffusionMBIR 38.20 0.966 38.22 0.966 38.19 0.967 37.86 0.965
SWORD 43.66 0.974 42.34 0.969 40.56 0.961 37.64 0.943
FBP × 23.94 0.548 20.27 0.494 17.46 0.445 14.68 0.380
NeRP × 27.03 0.809 27.12 0.814 26.81 0.799 26.86 0.806
R2-Gaussian × 42.58 0.982 41.93 0.979 41.56 0.975 41.03 0.972
GIFT × 43.73 0.985 42.91 0.984 42.15 0.982 40.55 0.977
Hepatic Cyst Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.74 0.930 35.52 0.905 31.33 0.791 33.36 0.854
MCG 37.87 0.891 37.91 0.891 37.94 0.891 37.94 0.891
DiffusionMBIR 37.92 0.955 38.02 0.957 37.91 0.956 37.50 0.952
SWORD 42.84 0.973 41.42 0.967 39.81 0.960 37.12 0.946
FBP × 25.26 0.603 19.94 0.525 17.27 0.475 14.26 0.416
NeRP × 26.65 0.808 26.57 0.804 26.65 0.808 26.39 0.799
R2-Gaussian × 42.05 0.976 41.57 0.972 40.83 0.967 40.02 0.966
GIFT × 42.96 0.981 42.29 0.980 41.47 0.977 39.12 0.971
Elbow Fracture Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 34.41 0.847 34.05 0.789 27.42 0.777 29.82 0.732
MCG 37.20 0.857 37.13 0.858 37.08 0.856 36.80 0.852
DiffusionMBIR 37.06 0.932 36.93 0.930 36.89 0.931 36.75 0.930
SWORD 42.83 0.959 38.67 0.917 37.39 0.901 34.71 0.865
FBP × 26.15 0.459 22.32 0.382 19.93 0.337 16.95 0.279
NeRP × 28.14 0.826 28.31 0.827 28.06 0.823 28.18 0.835
R2-Gaussian × 41.99 0.950 41.37 0.946 40.63 0.942 40.12 0.939
GIFT × 42.82 0.961 41.94 0.954 41.19 0.949 39.97 0.948
Spinal Fracture Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 23.86 0.866 23.94 0.841 23.92 0.832 23.70 0.810
MCG 38.52 0.913 38.52 0.913 38.48 0.912 38.40 0.911
DiffusionMBIR 39.34 0.973 39.27 0.973 39.08 0.972 38.49 0.969
SWORD 40.94 0.946 38.02 0.930 34.68 0.901 28.85 0.834
FBP × 16.41 0.793 15.20 0.766 14.73 0.741 13.96 0.698
NeRP × 28.10 0.847 26.24 0.779 27.95 0.840 26.48 0.790
R2-Gaussian × 40.51 0.970 39.06 0.961 38.33 0.957 37.98 0.953
GIFT × 41.23 0.981 39.70 0.977 38.41 0.971 37.68 0.968
Foot Fracture Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 37.53 0.860 35.61 0.783 32.52 0.817 32.46 0.837
MCG 39.40 0.891 39.62 0.895 39.43 0.894 39.45 0.894
DiffusionMBIR 40.45 0.956 40.31 0.955 40.22 0.954 40.26 0.957
SWORD 34.92 0.927 36.40 0.905 31.95 0.866 28.33 0.783
FBP × 23.45 0.235 18.80 0.181 17.23 0.160 14.46 0.132
NeRP × 30.69 0.921 30.82 0.926 30.76 0.932 30.56 0.927
R2-Gaussian × 39.97 0.960 39.12 0.956 38.46 0.951 37.90 0.949
GIFT × 41.81 0.981 41.21 0.980 40.51 0.974 39.60 0.977
Wrist Fracture Reconstruction
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 36.61 0.810 34.73 0.825 31.73 0.870 30.78 0.744
MCG 37.14 0.887 37.53 0.889 37.65 0.890 37.64 0.889
DiffusionMBIR 36.91 0.953 36.94 0.954 36.73 0.952 36.31 0.950
SWORD 36.74 0.903 33.91 0.874 31.57 0.832 28.91 0.766
FBP × 21.35 0.231 17.95 0.197 15.69 0.174 12.96 0.146
NeRP × 29.55 0.893 28.77 0.891 29.62 0.897 29.56 0.893
R2-Gaussian × 38.42 0.974 38.04 0.973 37.66 0.970 36.98 0.967
GIFT × 40.28 0.984 39.71 0.983 38.89 0.976 37.78 0.973
Subarachnoid Hemorrhage
Method Pretrain 180-view 120-view 90-view 60-view
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
RED-CNN 29.52 0.874 29.65 0.877 29.55 0.877 29.42 0.863
MCG 38.78 0.908 38.87 0.909 38.81 0.908 38.79 0.908
DiffusionMBIR 39.46 0.975 39.38 0.975 39.20 0.974 38.70 0.973
SWORD 42.54 0.965 39.60 0.955 36.71 0.938 31.84 0.895
FBP × 21.31 0.440 20.84 0.423 19.22 0.404 17.93 0.361
NeRP × 23.72 0.760 23.34 0.760 23.84 0.800 24.04 0.791
R2-Gaussian × 41.89 0.973 41.11 0.968 40.35 0.964 39.77 0.956
GIFT × 43.29 0.993 42.74 0.993 41.98 0.992 41.02 0.992