aramis-lab/miccai2022-stratifiad

MICCAI 2022 - Paper ID: 2116

How to cite the paper:

If you use this code and/or the dataset, please cite the following paper:

@InProceedings{miccai-paper-2116,
author="Jimenez, Gabriel and Kar, Anuradha and Ounissi, Mehdi and Ingrassia, Léa and Boluda, Susana and Delatour, Benoît and Stimmer, Lev and Racoceanu, Lev",
title="Visual DL-based explanation for neuritic plaques segmentation in Alzheimer's Disease",
booktitle="Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2022",
year="2022",
publisher="Springer International Publishing",
}

Setting up the OS:

• This code was develop and runs properly in Ubuntu 18.04 and with Python 3.6.9

sudo apt update

• In case pip3 is not installed, then run:

sudo apt install python3-pip

• Install pipenv:

pip3 install --user pipenv

Setting up the environment and dependencies:

• Inside the main directory, run the following command to syncronize all the dependencies:

pipenv sync

• To activate the virtual environment:

pipenv shell

Structure of the dataset:

The acquisition of the WSI and the patch generation process is described in the article. The dataset used for the three experiments below can be found here (size: 128x128) and here (size: 256x256). The structure is as follows:

dataset
├── 128x128
│   └── new_wsi_00
│   │   └── macenko
│   │   └── masks
│   │   └── patches
│   │   └── vahadane
│   └── new_wsi_01
│   │   └── macenko
│   │   └── masks
│   │   └── patches
│   │   └── vahadane
│   └── ...
└── 256x256
    └── new_wsi_00
    │   └── macenko
    │   └── masks
    │   └── patches
    └── new_wsi_01
    │   └── macenko
    │   └── masks
    │   └── patches
    └── ...

folder	description
macenko	contains all the patches and its corresponding corner augmentations using Macenko method for color normalization. Files are named in sequential order according to the annotations in the WSI and the augmented ones have an additional _Cx.png added to the name, where x is one of the four corners.
masks	contains all the masks and its corresponding corner augmentations. Files are named in sequential order according to the annotations in the WSI and the augmented ones have an additional _Cx.png added to the name, where x is one of the four corners.
patches	contains all the original unnormalized patches and its corresponding corner augmentations. Files are named in sequential order according to the annotations in the WSI and the augmented ones have an additional _Cx.png added to the name, where x is one of the four corners.
vahadane	contains all the patches and its corresponding corner augmentations using Vahadane method for color normalization. Files are named in sequential order according to the annotations in the WSI and the augmented ones have an additional _Cx.png added to the name, where x is one of the four corners. Only available for 128x128.

Running the experiments:

UNet - Experiment 01:

This experiment uses the 8 WSI images in the dataset and the Macenko color normalization method. We use this experiment to analyze how the network performs under different modalities and also when increasing the environment of the object (i.e., different patch sizes).

• For the 128x128 patch-size dataset, with cross validation and cross testing:

cd unet
sh test_00_128x128.sh
sh test_01_128x128.sh
sh test_02_128x128.sh
sh test_03_128x128.sh

• For the 256x256 patch-size dataset, with cross validation and cross testing:

cd unet
sh test_00_256x256.sh
sh test_01_256x256.sh
sh test_02_256x256.sh
sh test_03_256x256.sh

UNet - Experiment 02:

This experiment evaluates the impact of the scanner in the segmentation of plaques. We use only the 128x128 patch-size dataset as we obtained better performance with this patches in the first experiment.

• To train using the WSI from the scanner Hamamatsu NanoZoomer 2.0-RS:

cd unet
sh test_00_128x128_oldscan.sh

• To train using the WSI from the scanner Hamamatsu NanoZoomer S60:

cd unet
sh test_00_128x128_newscan.sh

UNet - Experiment 03:

This experiment evaluates the impact of the color normalization. We use only the 128x128 patch-size dataset and the configuration of the best fold from the first experiment.

cd unet
sh test_00_128x128_bestfold_vahadane.sh

Attention UNet:

This experiment evaluates the performance of the attention UNet in different patch size datasets. You will need to manually configure the following line in the file train_att_unet.py (provided that your data follows the structure mentioned above):

• To use the 128x128 dataset:

main_data_dir = os.path.join('..', '..', '..','dataset','128x128')

• To use the 256x256 dataset:

main_data_dir = os.path.join('..', '..', '..','dataset','256x256')

• To run the experiment:

python train_att_unet.py

Full results:

UNet and Attention UNet:

Table 01. Architecture used: UNet. Patch size: 128x128 pixels (best fold is reported in bold font).

fold_name	dev_dice	dev_f1	dev_recall	dev_precision	test_dice	test_f1	test_recall	test_precision
test_00_cv_00	0.7151	0.7165	0.6674	0.8017	0.6753	0.6707	0.668	0.784
test_00_cv_01	0.7034	0.6933	0.699	0.718	0.7046	0.7035	0.7495	0.7475
test_00_cv_02	0.6963	0.6932	0.6873	0.7339	0.6781	0.6765	0.7423	0.7094
test_01_cv_00	0.7011	0.7037	0.714	0.7239	0.7032	0.6962	0.6684	0.8052
test_01_cv_01	0.7231	0.7118	0.6763	0.7801	0.7248	0.7192	0.7105	0.8067
test_01_cv_02	0.72	0.7217	0.7519	0.7185	0.7141	0.7068	0.6811	0.8166
test_02_cv_00	0.709	0.7156	0.7195	0.7423	0.7027	0.7004	0.7043	0.7855
test_02_cv_01	0.7127	0.7195	0.7012	0.7618	0.6643	0.6608	0.6444	0.7996
test_02_cv_02	0.6807	0.6838	0.6862	0.7167	0.6306	0.6296	0.6634	0.7316
test_03_cv_00	0.6765	0.6813	0.6788	0.7183	0.6845	0.6855	0.8061	0.666
test_03_cv_01	0.6959	0.6967	0.6244	0.8167	0.6883	0.879	0.7981	0.6777
test_03_cv_02	0.6112	0.6008	0.61	0.6325	0.6521	0.6545	0.8018	0.6245
mean	0.6954	0.6948	0.6847	0.7387	0.6852	0.6986	0.7198	0.7462
std	0.0289	0.0313	0.0373	0.0462	0.0260	0.0596	0.0561	0.0615
max	0.7231	0.7217	0.7519	0.8167	0.7248	0.879	0.8061	0.8166
min	0.6112	0.6008	0.61	0.6325	0.6306	0.6296	0.6444	0.6245

Table 02. Architecture used: Attention UNet. Patch size: 128x128 pixels (best fold is reported in bold font).

fold_name	dev_dice	dev_f1	dev_recall	dev_precision	test_dice	test_f1	test_recall	test_precision
test_00_cv_00	0.7654	0.7654	0.7833	0.8000	0.6405	0.6405	0.5959	0.8177
test_00_cv_01	0.7216	0.7216	0.7293	0.7890	0.6734	0.6734	0.6387	0.8258
test_00_cv_02	0.7428	0.7422	0.7425	0.8108	0.6499	0.6499	0.6320	0.7883
test_01_cv_00	0.7646	0.7646	0.7876	0.7934	0.6851	0.6851	0.6582	0.7979
test_01_cv_01	0.7272	0.7272	0.7916	0.7299	0.7128	0.7122	0.7454	0.7522
test_01_cv_02	0.7335	0.7335	0.7593	0.7772	0.6909	0.6909	0.6836	0.7890
test_02_cv_00	0.7658	0.7658	0.7946	0.7900	0.7184	0.7184	0.7109	0.8060
test_02_cv_01	0.7159	0.7153	0.7565	0.7429	0.7263	0.7263	0.7950	0.7254
test_02_cv_02	0.7809	0.7809	0.8232	0.7855	0.6948	0.6948	0.7241	0.7488
test_03_cv_00	0.8193	0.8193	0.8180	0.8559	0.6959	0.6959	0.6839	0.8008
test_03_cv_01	0.6962	0.6962	0.6843	0.7908	0.7140	0.7140	0.7781	0.7281
test_03_cv_02	0.7857	0.7857	0.8297	0.7896	0.7024	0.7024	0.8435	0.6571
mean	0.7516	0.7515	0.7750	0.7879	0.6920	0.6920	0.7074	0.7698
std	0.0334	0.0335	0.0408	0.0301	0.0254	0.0254	0.0703	0.0469
max	0.8193	0.8193	0.8297	0.8559	0.7263	0.7263	0.8435	0.8258
min	0.6962	0.6962	0.6843	0.7299	0.6405	0.6405	0.5959	0.6571

Table 03. Architecture used: UNet. Patch size: 256x256 pixels (best fold is reported in bold font).

fold_name	dev_dice	dev_f1	dev_recall	dev_precision	test_dice	test_f1	test_recall	test_precision
test_00_cv_00	0.707	0.7095	0.6557	0.7962	0.6423	0.6363	0.5839	0.8071
test_00_cv_01	0.6509	0.6563	0.6744	0.6604	0.6783	0.6743	0.6563	0.7767
test_00_cv_02	0.6663	0.6657	0.6725	0.6826	0.6368	0.6322	0.6356	0.7269
test_01_cv_00	0.6939	0.7099	0.7041	0.7319	0.6963	0.6893	0.6632	0.7998
test_01_cv_01	0.6411	0.642	0.6018	0.7157	0.6856	0.6785	0.6468	0.8008
test_01_cv_02	0.6793	0.6829	0.6892	0.6979	0.6718	0.6654	0.6487	0.7732
test_02_cv_00	0.6893	0.7077	0.716	0.716	0.6395	0.6352	0.6104	0.766
test_02_cv_01	0.6911	0.698	0.6344	0.8027	0.621	0.6155	0.5748	0.7816
test_02_cv_02	0.6462	0.649	0.6593	0.6638	0.5949	0.595	0.6676	0.6185
test_03_cv_00	0.6432	0.6529	0.6323	0.7029	0.6507	0.65	0.7633	0.6242
test_03_cv_01	0.6747	0.6877	0.6581	0.7453	0.6507	0.6521	0.8146	0.5945
test_03_cv_02	0.5391	0.5355	0.5558	0.5535	0.5834	0.5856	0.8118	0.5149
mean	0.660	0.666	0.654	0.706	0.646	0.642	0.673	0.715
std	0.042	0.046	0.042	0.063	0.033	0.031	0.077	0.096
max	0.707	0.710	0.716	0.803	0.696	0.689	0.815	0.807
min	0.5391	0.5355	0.5558	0.5535	0.5834	0.5856	0.5748	0.5149

Table 04. Architecture used: Attention UNet. Patch size: 256x256 pixels (best fold is reported in bold font).

fold_name	dev_dice	dev_f1	dev_recall	dev_precision	test_dice	test_f1	test_recall	test_precision
test_00_cv_00	0.7288	0.7288	0.8318	0.6819	0.6439	0.6439	0.6312	0.7429
test_00_cv_01	0.6698	0.6698	0.7152	0.6855	0.6038	0.6038	0.5860	0.7462
test_00_cv_02	0.6644	0.6644	0.6396	0.7920	0.5638	0.5638	0.5132	0.7701
test_01_cv_00	0.7349	0.7349	0.8246	0.6971	0.6560	0.6560	0.6724	0.7162
test_01_cv_01	0.6471	0.6471	0.6618	0.7102	0.6796	0.6790	0.6746	0.7599
test_01_cv_02	0.6774	0.6774	0.6539	0.7842	0.6307	0.6307	0.5998	0.7847
test_02_cv_00	0.7463	0.7463	0.8221	0.7172	0.6323	0.6323	0.6289	0.7087
test_02_cv_01	0.6767	0.6761	0.7026	0.7199	0.6497	0.6497	0.6715	0.6983
test_02_cv_02	0.6883	0.6883	0.7973	0.6619	0.6194	0.6194	0.6652	0.6538
test_03_cv_00	0.7855	0.7855	0.7662	0.8413	0.6413	0.6413	0.6032	0.7821
test_03_cv_01	0.6199	0.6199	0.5635	0.8123	0.6731	0.6731	0.6997	0.7158
test_03_cv_02	0.6783	0.6783	0.8009	0.6438	0.6169	0.6169	0.8428	0.5326
mean	0.6931	0.6931	0.7316	0.7289	0.6342	0.6342	0.6490	0.7176
std	0.0447	0.0447	0.0846	0.0606	0.0301	0.0300	0.0762	0.0667
max	0.7855	0.7855	0.8318	0.8413	0.6796	0.6790	0.8428	0.7847
min	0.6199	0.6199	0.5635	0.6438	0.5638	0.5638	0.5132	0.5326

Scanner differences:

Table 05. Architecture used: UNet. Patch size: 128x128 pixels (best fold is reported in bold font). Scanner: Hamamatsu NanoZoomer 2.0-RS.

fold_name	dev_dice	dev_f1	dev_recall	dev_precision
test_00_cv_00	0.7361	0.7354	0.693	0.8106
test_00_cv_01	0.7234	0.7284	0.7451	0.734
test_00_cv_02	0.7388	0.7338	0.7155	0.7744
test_00_cv_03	0.7384	0.732	0.7063	0.7818
mean	0.7342	0.7324	0.7150	0.7752
std	0.0063	0.0026	0.0191	0.0274
max	0.7388	0.7354	0.7451	0.8106
min	0.7234	0.7284	0.693	0.734

Table 06. Architecture used: UNet. Patch size: 128x128 pixels (best fold is reported in bold font). Scanner: Hamamatsu NanoZoomer S60.

fold_name	dev_dice	dev_f1	dev_recall	dev_precision
test_00_cv_00	0.6286	0.6429	0.6121	0.7143
test_00_cv_01	0.6757	0.6695	0.6931	0.6855
test_00_cv_02	0.617	0.6448	0.6668	0.6443
test_00_cv_03	0.6167	0.6445	0.6653	0.6452
mean	0.6345	0.6504	0.6593	0.6723
std	0.0243	0.0110	0.0294	0.0294
max	0.6757	0.6695	0.6931	0.7143
min	0.6167	0.6429	0.6121	0.6443