Deep learning using computed tomography to identify high-risk patients for acute small bowel obstruction: development and validation of a prediction model

PyTorch implementation and pretrained models for DBA+DRP. For details, see the paper: Deep learning using computed tomography to identify high-risk patients for acute small bowel obstruction: development and validation of a prediction model

Pre-trained models

You can downlaod weights of the backbones used in our experiments. Detailed arguments and learning/assessment methods can be found in the configs folder. The path of weights is in the top-level of the project folder(./intestinal_obstruction/weights).

Backbone	Methods	Accuracy(%)	AUROC	Download
ResNet	Naive	66.15	0.848 + 0.05	Link
	DBA+DRP	70.26	0.876 + 0.02	Link
ResNext	Naive	68.46	0.874 + 0.02	Link
	DBA+DRP	66.41	0.883 + 0.01	Link
WideResNet	Naive	65.13	0.861 + 0.02	Link
	DBA+DRP	72.56	0.896 + 0.01	Link
DenseNet	Naive	75.12	0.868 + 0.05	Link
	DBA+DRP	72.68	0.873 + 0.07	Link
EfficientNet	Naive	71.22	0.841 + 0.04	Link
	DBA+DRP	71.71	0.868 + 0.03	Link

Training

Requriements

Please install the requirements including pytorch for stable running. This code has been developed with python3.8 and PyTorch 2.0.1, torch 0.19, and CUDA 11.8.

pip install -r requirements.txt

Single training

Run on general train/valid dataset with single GPU for 100 epochs as following command. You can choose specific backbone and methods using - notation(backbone=resnet or backbone=resnet-dbadrp). For other configs and config structures, please check out the configs folder.

python main.py dataset=asbo gpus=[0] train.epochs=100 backbone=resnet-dbadrp

K-Fold training

By default, we train using 5-fold cross validation and evaluated the model with average values. For 5-fold training, all you need to do is change a few commands for the dataset and execute file.

python cross_validation_train.py dataset=asbo-k gpus=[0] backbone=resnet-dbadrp

Inference

Before checking the model performance, the all weight files should be located in weights forder in the foramt ‘model_{n}k_best.pt’(resnet0k_best.pt, resnet1k_best.pt, …., etc.). If weight files are located well, you just run. This will automatically evaluate about all 5 folds.

python cross_inference.py

Visualization of featuremap

To generate the GradCAM of last layer, please run the following command. However, as with inference, the model must be in the right directory before running the command. You can modify the method, gpus, k, etc.

python gradcam.py

Folder structure of dataset

The data has three classes, HGSBO, LGSBO and NORMAL, which are located in the top-level folder. The three classes have folders separated by case series id, and each folder has a .dcm file containing CT images. The data splitting according to k-fold is stored in the splits folder as a .json file, and the corrupted data for the robustness test is organized under the distort folder with distory_type, intenstiy, and classes.

./data
├── HGSBO
│   └── case_series_id
|		└── 0001.dcm
|		└── ...
├── LGSBO
│   └── case_series_id
|		└── 0001.dcm
|		└── ...
├── NORMAL
│   └── case_series_id
|		└── 0001.dcm
|		└── ...
├── splits
|    └── asbo_k_classification.json
|    └── asbo_classification.json
└──distort
   └──affine # dist_type
       └── 0 # intensity
           ├── HGSBO #class
           ├── LGSBO #class
           └── NORMAL #class

The *k*.json file is constructed by list with 5 length, and each item of list is dict with keys: train, train_label, valid, valid_label. On the other hand, a .json file without k is just a dict and has the same keys.

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Citation

If our project is helpful for your research, please consider citation:paperclip: and giving a star:star2: :

@article{oh2023deep,
  title={Deep learning using computed tomography to identify high-risk patients for acute small bowel obstruction: development and validation of a prediction model: A retrospective cohort study},
  author={Oh, Seungmin and Ryu, Jongbin and Shin, Ho-Jung and Song, Jeong Ho and Son, Sang-Yong and Hur, Hoon and Han, Sang-Uk},
  journal={International Journal of Surgery},
  pages={10--1097},
  year={2023},
  publisher={LWW}
}