SGDA

This repository includes codes, models, and test results for our paper: "Towards 3D Universal Pulmonary Nodule Detection via Slice Grouped Domain Attention". This project is licensed for non-commerical research purpose only.

Results and Models

Comparison of our SGDA and other multi-domain methods in terms of FROC on dataset LUNA16, tianchi, and russia. Values below the names of datasets are FROCs (unit: %). All the methods utilize NoduleNet as backbone: (1) shared models with the prefix 'uni-', (2) independent models with the word 'single' in the name, (3) multi-domain methods, (4) universal models with 'SG' in the name (Ours).

Method	#Adapters	#Groups	#Params	LUNA16	tianchi	russia	Avg	Pre-trained Model
single NoduleNet	-	-	16.73Mx3	77.71	68.23	37.19	61.04	model_luna16 model_tianchi model_russia & res_luna16 res_tianchi res_russia
uniNoduleNet	-	-	39.50M	79.88	68.60	33.35	60.61	model & res_luna16 res_tianchi res_russia
NoduleNet+BN	3	-	39.51M	79.94	68.12	36.52	61.52	model & res_luna16 res_tianchi res_russia
NoduleNet+series	3	-	40.14M	78.44	70.41	33.39	60.74	model & res_luna16 res_tianchi res_russia
NoduleNet+parallel	3	-	40.13M	78.57	70.14	35.61	61.44	model & res_luna16 res_tianchi res_russia
NoduleNet+separable	3	-	34.68M	66.31	62.26	32.96	53.84	model & res_luna16 res_tianchi res_russia
NoduleNet+SNR	-	-	39.50M	69.52	66.57	36.76	57.61	model & res_luna16 res_tianchi res_russia
single NoduleNet+SE	-	-	16.74Mx3	77.78	68.86	38.06	61.56	model_luna16 model_tianchi model_russia & res_luna16 res_tianchi res_russia
uniSENoduleNet	-	-	39.51M	80.53	69.13	34.34	61.33	model & res_luna16 res_tianchi res_russia
NoduleNet+SE	3	-	39.54M	78.89	72.33	35.89	62.37	model & res_luna16 res_tianchi res_russia
DANoduleNet	3	-	39.54M	82.63	73.29	38.50	64.80	model & res_luna16 res_tianchi res_russia
single NoduleNet+SGSE	-	4	16.77Mx3	78.30	70.36	39.01	62.55	model_luna16 model_tianchi model_russia & res_luna16 res_tianchi res_russia
uniSGSENoduleNet	-	4	39.54M	81.12	71.00	38.42	63.51	model & res_luna16 res_tianchi res_russia
NoduleNet+SGSE	3	4	39.62M	80.93	70.94	38.30	63.39	model & res_luna16 res_tianchi res_russia
SGDANoduleNet	3	4	39.82M	81.91	77.13	37.15	65.39	model & res_luna16 res_tianchi res_russia

Comparison of our SGDA and other multi-domain methods in terms of FROC on dataset PN9. The values are pulmonary nodule detection sensitivities (unit: %) with each column representing the average number of false positives per CT image. All the methods utilizes SANet as backbone: (1) baseline model with the prefix 'uni-', (2) universal models with 'SG' in the name (Ours).

Method	#Adapters	#Groups	#Params	0.125	0.25	0.5	1.0	2.0	4.0	8.0	Avg	Pre-trained Model
uniSANet	-	-	15.28M	38.08	45.05	54.46	64.50	75.33	83.86	89.96	64.46	model & res
DASANet	3	-	15.32M	54.86	54.86	54.86	64.94	75.43	83.53	88.18	68.09	model & res
*SGDASANet w/o CA	3	4	15.36M	52.06	52.06	58.63	66.33	77.05	85.13	90.12	68.77	model & res
*SGDASANet w/ CA	3	4	15.45M	57.63	57.63	57.63	65.73	75.09	83.56	88.25	69.36	model & res

Requirements

The code is built with the following libraries:

• Python 3.6 or higher
• CUDA 10.0 or higher
• PyTorch 1.2 or higher
• tqdm
• scipy

Besides, you need to install a custom module for bounding box NMS and overlap calculation.

cd build/box
python setup.py install

Data

Pulmonary nodule datasets. 'Scans' denotes the number of CT scans. 'Nodules' denotes the number of labeled nodules. 'Class' denotes the class number. And 'Raw' means whether the dataset contains raw CT scans. 'Image Size' gives the dimensions of the CT image matrix alont the x, y, and z axes. 'Spacing' gives the voxel sizes (mm) along the x, y, and z axes.

Dataset	Year	Scans	Nodules	Class	Raw	File Size	Image Size	Spacing	Source Link
LUNA16	2016	601	1186	2	Yes	25M-258M	512x512x95-512x512x733	(0.86,0.86,2.50)-(0.64,0.64,0.50)	link & split
tianchi	2017	800	1244	2	Yes	26M-343M	512x512x114-512x512x1034	(0.66,0.66,2.50)-(0.69,0.69,0.30)	link & split
russia	2018	364	1850	2	Yes	80M-491M	512x512x313-512x512x1636	(0.62,0.62,0.80)-(0.78,0.78,0.40)	link & split
PN9	2021	8796	40436	9	No	5.6M-73M	212x212x181-455x455x744	(1.00,1.00,1.00)-(1.00,1.00,1.00)	link & split

Download the datasets and add the information to configs/*config*.py. Please refer to specificFiles/LIDC/lung_seg.py and specificFiles/LIDC/preprocess.py for the data preprocessing.

Testing

Run the following scripts to evaluate the model and obtain the results of FROC analysis.

python universal_test_sanet.py --ckpt='./results/model/model.ckpt' --save_dir='./results/'

Training

This implementation supports multi-gpu, data_parallel training.

Change training configuration and data configuration in configs/*config*.py, especially the path to preprocessed data.

Run the training script:

python SGDA_train_sanet_middle_top.py

Citations

If you are using the code/model/data provided here in a publication, please consider citing:

@article{SGDA22,  
author={Rui Xu and Zhi Liu and Yong Luo and Han Hu and Li Shen and Bo Du and Kaiming Kuang and and Jiancheng Yang},   
title={Towards 3D Universal Pulmonary Nodule Detection via Slice Grouped Domain Attention},   
journal={},    
year={},  
volume={},  
number={},  
pages={},  
doi={}} 

Contact

For any questions, please contact: rui.xu AT whu.edu.cn.

Acknowledgment

This code is based on the UOD, SANet and NoduleNet.