🆕News | 🛠️Installation | 👀Model Zoo | 🤔Reporting Issues

🌻 This is an official implementation for paper MedISeg

🌻 Here is a brief introduction on 知乎

Introduction

In this paper, we collect a series of MedISeg tricks for different model implementation phases (i.e., pre-training model, data pre-processing, data augmentation, model implementation, model inference, and result post-processing), and experimentally explore the effectiveness of these tricks on the consistent baseline models. Compared to paper-driven surveys that only blandly focus on the advantages and limitation analyses of segmentation models, our work provides a large number of solid experiments and is more technically operable. With the extensive experimental results on both the representative 2D and 3D medical image datasets, we explicitly clarify the effect of these tricks.

The surveyed medical image segmentation tricks and their latent relations

Citation

🌻 If you use this toolbox or benchmark in your research, please cite:

@article{zhang2022deep,
  title={Deep Learning for Medical Image Segmentation: Tricks, Challenges and Future Directions},
  author={Zhang, Dong and Lin, Yi and Chen, Hao and Tian, Zhuotao and Yang, Xin and Tang, Jinhui and Cheng, Kwang Ting},
  journal={arXiv preprint arXiv:2209.10307},
  year={2022}
}

News

🌻 1.1.0 was released in 01/09/2022

• Supported Backbones:

• ResNet (CVPR'2016)

• DenseNet (CVPR'2017)

• ViT (ICLR'2021)

• Supported Methods:

• 2D-UNet (MICCAI'2015)

• 3D-UNet (MICCAI'2016)

• Supported Datasets:

• ISIC 2018

• CoNIC

• KiTS19

• LiTS

Installation

• Option 1:

pip install -r requirements.txt

• Option 2:

pip install albumentations
pip install ml_collections
pip install numpy 
pip install opencv-python
pip install pandas
pip install rich
pip install SimpleITK
pip install timm
pip install torch
pip install tqdm
pip install nibabel
pip install medpy

Data Preparation

Please download datasets from the official website:

• ISIC 2018: 2D ISIC 2018 Lesion Boundary Segmentation Dataset
• CoNIC: 2D Colon Nuclei Identification and Counting Challenge Dataset
• KiTS19: 3D Kidney Tumor Segmentation 2019 Dataset
• LiTS17: 3D Liver Tumor Segmentation 2017 Dataset

Inference with Pre-trained Models

Download the trained weights from Model Zoo.

Run the following command for 2DUNet:

python 2DUNet/NetworkTrainer/test.py --test-model-path $YOUR_MODEL_PATH

Run the following command for 3DUNet:

python 3DUNet/NetworkTrainer/test.py --test-model-path $YOUR_MODEL_PATH

Training & Evaluation

We provide the shell scripts for training and evaluation by 5-fold cross-validation.

Run the following command for 2DUNet:

sh 2DUNet/config/baseline.sh

Run the following command for 3DUNet:

sh 3DUNet/config/baseline.sh

And the commands train/test with various tricks are also provided in */config/. For the details of the segmentation tricks, please refer to the paper.

Visualization

From top to bottom: raw image, ground truth, prediction.

ISIC 2018

CoNIC

KiTS19

LiTS17

Model Zoo

• Since our Google space is very limited, here we only provide a part of the weight links.

• In each 5-fold cross-validation, here we only release a weight with a higher performance.

• The full weights can be downloaded from Baidu Netdisk.

Training weights on ISIC 2018:

Dataset	Baseline	Method	Recall (%)	Percision (%)	Dice (%)	IoU (%)	Weight
ISIC 2018	2DUNet	PyTorch	88.37%	91.41%	88.17%	86.66%	weight
ISIC 2018	2DUNet	+ Image-21K	90.06%	92.64%	90.07%	88.44%	weight
ISIC 2018	2DUNet	+ GTAug-B	88.46%	93.22%	89.19%	87.68%	weight
ISIC 2018	2DUNet	+ CBL(Tvers)	89.93%	90.47%	88.53%	86.72%	weight
ISIC 2018	2DUNet	+ TTAGTAug-B	89.74%	92.40%	89.61%	88.14%	-
ISIC 2018	2DUNet	+ EnsAvg	90.80%	90.88%	89.32%	87.72%	weight

Training weights on CoNIC:

Dataset	Baseline	Method	Recall (%)	Percision (%)	Dice (%)	IoU (%)	Weight
CoNIC	2DUNet	PyTorch	77.76%	74.71%	75.76%	77.17%	weight
CoNIC	2DUNet	+ Image-21K	80.59%	76.71%	78.25%	79.14%	weight
CoNIC	2DUNet	+ GTAug-B	81.23%	80.57%	80.53%	81.02%	weight
CoNIC	2DUNet	+ TTAGTAug-A	80.22%	79.29%	79.28%	79.98%	-

Training weights on KiTS19:

Dataset	Baseline	Method	Recall (%)	Percision (%)	Dice (%)	IoU (%)	Weight
KiTS19	3DUNet	PyTorch	93.69%	95.28%	94.32%	89.44%	weight
KiTS19	3DUNet	+ EnsAvg	94.46%	96.29%	95.27%	91.09%	weight

Training weights on LiTS17:

Dataset	Baseline	Method	Recall (%)	Percision (%)	Dice (%)	IoU (%)	Weight
LiTS17	3DUNet	PyTorch	93.66%	82.08%	87.00%	77.37%	weight
LiTS17	3DUNet	+ ModelGe	92.98%	80.80%	85.89%	75.63%	weight
LiTS17	3DUNet	Patching192	95.33%	94.67%	94.87%	90.40%	weight
LiTS17	3DUNet	+ GTAug-A	92.08%	73.40%	81.15%	68.71%	weight
LiTS17	3DUNet	+ OHEM	92.50%	82.78%	86.81%	77.12%	weight
LiTS17	3DUNet	+ EnsAvg	92.10%	87.21%	89.07%	80.70%	weight
LiTS17	3DUNet	+ ABL-CS	93.65%	84.97%	88.60%	80.01%	-

Todo list

• Experiments on more datasets

• Experiments on other backbones

• Experiments on more tricks

• Other interesting attempts

Announcements

🌻 We welcome more like-minded friends to join in this project and continue to expand this storage

🌻 If you have any suggestions or comments please let us know

🌻 If you have any problems in using this code, please contact: [email protected] or [email protected]