K-Net/contrib/MedicalSeg/configs/synapse

• Multi-Atlas Labeling
• • Abdomen Dataset
  • • CT Scans
    • Organs to Segment
    • Prepare dataset
    • Inference helper User Guide
  • Performance
  • • TransUnet
    • SwinUnet

Multi-Atlas Labeling

Multi-atlas labeling has proven to be an effective paradigm for creating segmentation algorithms from training data. These approaches have been extraordinarily successful for brain and cranial structures (e.g., our prior MICCAI workshops: MLSF’11, MAL’12, SATA’13). After the original challenges closed, the data continue to drive scientific innovation; 144 groups have registered for the 2012 challenge (brain only) and 115 groups for the 2013 challenge (brain/heart/canine leg). However, innovation in application outside of the head and to soft tissues has been more limited. This workshop will provide a snapshot of the current progress in the field through extended discussions and provide researchers an opportunity to characterize their methods on a newly created and released standardized dataset of abdominal anatomy on clinically acquired CT. The datasets will be freely available both during and after the challenge.

Abdomen Dataset

CT Scans

Under Institutional Review Board (IRB) supervision, 50 abdomen CT scans of were randomly selected from a combination of an ongoing colorectal cancer chemotherapy trial, and a retrospective ventral hernia study. The 50 scans were captured during portal venous contrast phase with variable volume sizes (512 x 512 x 85 - 512 x 512 x 198) and field of views (approx. 280 x 280 x 280 mm3 - 500 x 500 x 650 mm3). The in-plane resolution varies from 0.54 x 0.54 mm2 to 0.98 x 0.98 mm2, while the slice thickness ranges from 2.5 mm to 5.0 mm. The standard registration data was generated by NiftyReg.

Organs to Segment

Thirteen abdominal organs were manually labeled by two experienced undergraduate students, and verified by a radiologist on a volumetric basis using the MIPAV software, including:

1. spleen
2. right kidney
3. left kidney
4. gallbladder
5. esophagus
6. liver
7. stomach
8. aorta
9. inferior vena cava
10. portal vein and splenic vein
11. pancreas
12. right adrenal gland
13. left adrenal gland

In the TransUnet model, the labels of the data are reduced to 8, which are:

1. spleen
2. right kidney
3. left kidney
4. gallbladder
5. liver
6. stomach
7. aorta
8. pancreas

Prepare dataset

To preprocess the synapse data, you first need to download RawData.zip from https://www.synapse.org/#!Synapse:syn3193805/files/. Then put it in the MedicalSeg/data/abdomen directory. Then run the tools/prepare_abdomen.py program,

mkdir data/abdomen
cp path/to/RawData.zip data/abdomen
python tools/prepare_abdomen.py

the dataset will be automatically generated. The file structure is as follows:

abdomen
|--RawData.zip
|--abdomen_raw
│   ├── RawData
│   │   ├──RawData
│   │   │   ├── Training
│   │   │   │   ├── img
│   │   │   │   │   ├── img0001.nii.gz
│   │   │   │   │   └── ...
│   │   │   │   └── ...
│   │   │   │   ├── label
│   │   │   │   │   ├── img0001.nii.gz
│   │   │   │   │   └── ...
│   │   │   │   └── ...
├── abdomen_phase0
│   ├── images
│   │   ├── img0001-0001.npy
│   │   └── ...
│   ├── labels
│   │   ├── label0001-0001.npy
│   │   └── ...
│   ├── train_list.txt
│   └── val_list.txt

In the prepare_abdomen.py program, the default training set and validation set split ratio is 6:4. If you want to modify the split ratio, you can modify line 113 of the program and pass in the train_split parameter. For example, 8:2, the code is as follows:

self.train_val_split(train_split=0.8)

Then you can start the training program, such as the following command for TransUnet:

python train.py --config configs/synapse/transunet_abdomen_224_224_1_14k_1e-2.yml --do_eval --save_interval 1000 --has_dataset_json False --is_save_data False --num_workers 4 --log_iters 10 --use_vdl

Inference helper User Guide

1. Since the shape of the input and output data of the TransUnet and SwinUnet networks is different from other networks, it is necessary to define special pre-process and post-process rules for these networks. This is achieved by InferenceHelper.
2. InferenceHelper is an abstract base class that contains two methods, preprocess and postprocess. If you need to add a new inference helper to your own network, you need to customize the class in the medicalseg/inference_helpers package and inherit the base InferenceHelper class.

class InferenceHelper2D(InferenceHelper):

3. MedicalSeg maintains an INFERENCE_HELPERS variable of type ComponentManager. You can add your custom inference helper through the add_component method.

@manager.INFERENCE_HELPERS.add_component
class InferenceHelper2D(InferenceHelper):

Also you need to import your class in the __init__.py file of the inference_helper package.

# in medicalseg/inference_helpers/__init__.py file

from .inference_helper_2d import InferenceHelper2D

4. You also need to implement preprocess and postprocess methods suitable for your own network, such as the following:

    def preprocess(self, cfg, imgs_path, batch_size, batch_id):
        for img in imgs_path[batch_id:batch_id + batch_size]:
            im_list = []
            imgs = np.load(img)
            imgs = imgs[:, np.newaxis, :, :]
            for i in range(imgs.shape[0]):
                im = imgs[i]
                im = cfg.transforms(im)[0]
                im_list.append(im)
            img = np.concatenate(im_list)
        return img

    def postprocess(self, results):
        results = np.argmax(results, axis=1)
        results = results[np.newaxis, :, :, :, :]
        return results

5. Finally you can reference your inference helper in your yaml configuration file. This way, the customized inference helper will be automatically constructed and invoked in the inference script.

# in yml file of configs

export:
  inference_helper:
    type: InferenceHelper2D

Performance

TransUnet

Chen, Jieneng and Lu, Yongyi and Yu, Qihang and Luo, Xiangde and Adeli, Ehsan and Wang, Yan and Lu, Le and Yuille, Alan L., and Zhou, Yuyin. "TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation." arXiv preprint arXiv:2102.04306, 2021.

Backbone	Resolution	lr	Training Iters	Dice	Links
R50-ViT-B_16	224x224	1e-2	13950	81.05%	model | log | vdl

SwinUnet

Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, Manning Wang. "Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation." arXiv preprint arXiv:2105.05537, 2021.

Backbone	Resolution	lr	Training Iters	Dice	Links
SwinTransformer-tinier	224x224	5e-2	14000	82.062%	model | log | vdl