Skip to content

Latest commit

 

History

History
195 lines (140 loc) · 13.8 KB

AMOS.md

File metadata and controls

195 lines (140 loc) · 13.8 KB

AMOS 2022

Dataset Information

Abdominal multi-organ segmentation is one of the most attractive topics in the field of medical image analysis, playing an important role in supporting clinical workflows such as disease diagnosis and treatment. The recent success of deep learning methods applied to abdominal multi-organ segmentation has exposed the lack of large-scale benchmarks to develop and compare these methods. Although there are several benchmark datasets for abdominal organ segmentation, the limited number of organs of interest and training samples still restrict the capabilities of modern deep models and make it difficult to provide a comprehensive and fair estimate of different methods.

Moreover, most current medical image analysis research focuses on building customized systems to handle monotonous inputs and outputs related to a single task. The complexity of such systems increases dramatically when the inputs or outputs become more diversified. If an algorithm can handle a variety of input patterns and outputs, then actual deployment will be greatly simplified, and it will align with human capabilities: the ability to absorb data from many sources, learn general rules, seamlessly integrate them, and serve a variety of tasks.

To address the aforementioned drawbacks and further promote the development of medical image segmentation technology, a large-scale, clinically-oriented, and diversified abdominal multi-organ segmentation benchmark dataset, AMOS, has been proposed. It provides 500 CT and 100 MR scans from multi-centers, multi-vendors, multi-modalities, multi-phases, and multi-disease patients, each with voxel-level annotations for 15 abdominal organs.

Compared to previous medical image analysis datasets, AMOS ensures a large data volume while providing fine annotations for up to 15 types of abdominal organs, possessing high clinical value.

Task information

Task 1 - Abdominal Organ Segmentation (CT only)

Task 1 is a more routine task, aimed at comprehensively evaluating the performance of different segmentation methods across a wide range and diversity of CT scans. It involves 500 patients and 15 organs (spleen, right kidney, left kidney, gallbladder, esophagus, liver, stomach, aorta, inferior vena cava, pancreas, right adrenal gland, left adrenal gland, duodenum, bladder, prostate/uterus).

Task 2 - Abdominal Organ Segmentation (CT and MR)

This task extends the imaging modality targets of Task 1 to MR modality. In this cross-modality setting, an algorithm is needed to segment abdominal organs from both CT and MR. Specifically, an additional 100 MR scans with the same type of annotations will be provided.

In traditional multi-modality segmentation tasks common in medical imaging, there are multiple modal inputs for each inference, for example, the BraTS dataset provides 4 types of MR modality inputs for each sample. Compared to the traditional setting, the cross-modality setting of AMOS is more challenging; the model will only input one modality at a time during training and inference, either CT or MR, requiring the model to adapt to both modalities."

Dataset Meta Information

Dimensions Modality Task Type Anatomical Structures Anatomical Area Number of Categories Data Volume File Format
3D CT, MRI Segmentation Spleen, right kidney, left kidney, gallbladder, esophagus, liver, stomach, aorta, inferior vena cava, pancreas, right adrenal gland, left adrenal gland, duodenum, bladder, prostate/uterus Abdomen 15 240 for training, 120 for validation, 240 for test .nii.gz

Resolution Details

Dataset Statistics spacing (mm) size
min (0.45, 0.45, 0.82) (192, 60, 64)
median (0.78, 0.78, 5) (512, 512, 100)
max (1.95, 1.95, 5) (768, 768, 512)

Intensity Information Statistics

CT Intensity Information Statistics

Organs CT - min CT - 0.5% CT - median CT - mean CT - std CT - 99.5% CT - max
Liver -90.03 -21.97 49.99 47.95 18.58 86 112
Right Kidney -75.99 -31 25.03 23.96 16 62.02 91.98
Spleen -103.03 -33.02 28.02 26.94 16.2 62.99 99
Pancreas -80.99 -31 20.02 19.79 15.38 57.99 122.99
Aorta -56.03 -31 39.98 36.83 19.98 80.45 96.01
Inferior Vena Cava -156.01 -12.02 57.99 56.87 17.93 98.03 338.02
Right Adrenal Gland -714.97 -221.36 26.98 22.03 39.45 75.99 176.03
Left Adrenal Gland -130 -38.02 42 39.93 20.91 86.98 315.99
Gallbladder -86.97 -39 42 39.63 21.98 91.98 122.99
Esophagus -91.98 -31 36.01 34.35 20.47 81.97 115.97
Stomach -75.01 -54.38 18.98 16.91 26.27 73 91.01
Duodenum -72.02 -43.03 25.03 23.15 21.97 72.03 85.03
Left Kidney -646 -98.12 34.98 31.41 28.92 81.97 166.02
Bladder -116.03 -33.02 13 12.52 12.79 44.99 502.02
Prostate/Uterus -57.01 -23.01 42 42.65 25.47 112.98 154.97

MR Intensity Information Statistics

Organs MR - min MR - 0.5% MR - median MR - mean MR - std MR - 99.5% MR - max
Liver 0 86 169 170.95 26.52 277 409
Right Kidney 0 38 166 166.14 38.39 272 364
Spleen 0 14 155 155.34 42.12 291 384
Pancreas 49 72.98 237.5 245.64 75.55 440.76 469
Aorta 0 0 141 131.62 48.39 224.55 247
Inferior Vena Cava 0 92 213 216.82 43.13 352 421
Right Adrenal Gland 0 0 168 153.38 62.73 277 374
Left Adrenal Gland 5 83 215 227.39 78.59 411.19 460
Gallbladder 25 61 143 144.95 34.98 270 392
Esophagus 4 97 258 252.16 46.85 365 427
Stomach 0 19.43 177.5 169.28 54.34 263.05 276
Duodenum 82 94.8 191.5 189.41 32.36 254.67 258
Left Kidney 0 0 167 155.1 64.18 285 350
Bladder nan nan nan nan nan nan nan
Prostate/Uterus nan nan nan nan nan nan nan

Label Information Statistics

Organ/Tissue Voxel Ratio (min) Voxel Ratio (median) Voxel Ratio (mean) Voxel Ratio (max) Actual Size cm³ (min) Actual Size cm³ (median) Actual Size cm³ (mean) Actual Size cm³ (max) Number of Instances
Spleen 1.15% 3.17% 3.42% 10.40% 1379 2571 2669 7330 200
Right Kidney 0.22% 0.17% 0.18% 0.52% 14 135 142 343 200
Left Kidney 0.17% 0.17% 0.19% 1.32% 10 142 150 1008 199
Gall Bladder 0.03% 0.03% 0.04% 0.29% 1 24 30 143 189
Esophagus 0.02% 0.02% 0.02% 0.07% 0 12 15 60 199
Liver 0.58% 1.52% 1.67% 5.08% 653 1231 1318 5840 200
Stomach 0.00% 0.31% 0.44% 2.87% 0 244 325 1254 198
Aorta 0.03% 0.12% 0.14% 0.57% 26 92 116 373 200
Inferior Vena Cava 0.02% 0.08% 0.09% 0.23% 31 69 69 135 200
Pancreas 0.02% 0.08% 0.09% 0.26% 15 70 69 183 200
Right Adrenal Gland 0.00% 0.00% 0.00% 0.01% 1 3 3 8 199
Left Adrenal Gland 0.00% 0.00% 0.01% 0.02% 0 3 3 9 200
Duodenum 0.02% 0.06% 0.07% 0.19% 19 50 55 146 200
Bladder 0.01% 0.11% 0.17% 2.98% 14 85 128 1255 195
Prostate/Uterus 0.01% 0.04% 0.07% 0.59% 7 30 51 268 192

Visualization

CT image example.

MR image example.

File Structure

After decompressing the official compressed package, the file structure is as follows, which conforms to the nnUNet standard format.

AMOS22
  ├── dataset.json
  ├── imagesTr
  │   ├── amos_0001.nii.gz
  │   ├── amos_0004.nii.gz
  │   ├── amos_0005.nii.gz
  │   ├── ...
  ├── labelsTr
  │   ├── amos_0001.nii.gz
  │   ├── amos_0004.nii.gz
  │   ├── amos_0005.nii.gz
  │   ├── ...
  ├── imagesTs
  │   ├── amos_0002.nii.gz
  │   ├── amos_0003.nii.gz
  │   ├── ...

Authors and Institutions

Xiang Wan (SRIBD)

Ruimao Zhang (CUHKSZ)

Zhen Li (CUHKSZ)

Guanbin Li (SYSU)

Ping Luo (HKU)

Yuanfeng Ji (HKU)

Weijie Ma (CUHKSZ)

Haotian Bai (CUHKSZ)

Jie Yang (CUHKSZ)

Chongjian Ge (HKU)

Lingyan Zhang (LGCHSZ)

Wanling Ma (LGDPHSZ)

Shaohua Zhou (ICTCAS & CUHKSZ)

Shuguang Cui (CUHKSZ)

Source Information

Official Website: https://amos22.grand-challenge.org/

Download Link: https://amos22.grand-challenge.org/Instructions/

Article Address: https://arxiv.org/abs/2206.08023

Release Date: 2022-05

Citation

@article{ji2022amos,
  title={AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation},
  author={Ji, Yuanfeng and Bai, Haotian and Yang, Jie and Ge, Chongjian and Zhu, Ye and Zhang, Ruimao and Li, Zhen and Zhang, Lingyan and Ma, Wanling and Wan, Xiang and others},
  journal={arXiv preprint arXiv:2206.08023},
  year={2022}
}

Original introduction article is here.