mi2rl/SMART-Net

It's scheduled to be uploaded soon.

SMART-Net - Official Pytorch Implementation

We proposed a supervised multi-task aiding representation transfer learning network called SMART-Net.

💡 Highlights

• Develop a robust feature extractor of brain hemorrhage in head & neck NCCT through three kinds of multi-task representation learning.
• Propose the consistency loss to alleviate the disparity of two pretext tasks' heads, resulting in improved transferability and representation. Connect the feature extractor with the target-specific 3D operator via transfer learning to expand volume-level tasks.
• Explore relationships of the proposed multi-pretext task combinations and perform ablation studies on optimal 3D operators for volume-level ICH tasks.
• Validate the model on multiple datasets with previous methods and ablation studies for the robustness and practicality of our method.

Paper

This repository provides the official implementation of training SMART-Net as well as the usage of the pre-trained SMART-Net in the following paper:

Improved performance and robustness of multi-task representation learning with consistency loss between pretexts for intracranial hemorrhage identification in head CT
Sunggu Kyung¹, Keewon Shin, Hyunsu Jeong, Ki Duk Kim, Jooyoung Park, Kyungjin Cho, Jeong Hyun Lee, Gil-Sun Hong, and Namkug Kim
MI2RL LAB
(Under revision...) Medical Image Analysis (MedIA)

code

Requirements

• Linux
• Python 3.8.5
• PyTorch 1.8.0

📦 SMART-Net Framework

1. Clone the repository and install dependencies

$ git clone https://github.com/babbu3682/SMART-Net.git
$ cd SMART-Net/
$ pip install -r requirements.txt

2. Preparing data

For your convenience, we have provided few 3D nii samples from Physionet publish dataset as well as their mask labels.

Note: We do not use this data as a train, it is just for code publishing examples.

You can use your own data using the dicom2nifti for converting from dicom to nii.

• The processed hemorrhage directory structure

datasets/samples/
    train
        |--  sample1_hemo_img.nii.gz
        |--  sample1_hemo_mask.nii
        |--  sample2_normal_img.nii.gz
        |--  sample2_normal_mask.nii        
                .
                .
                .
    valid
        |--  sample9_hemo_img.nii.gz
        |--  sample9_hemo_mask.nii
        |--  sample10_normal_img.nii.gz
        |--  sample10_normal_mask.nii
                .
                .
                .
    test
        |--  sample20_hemo_img.nii.gz
        |--  sample20_hemo_mask.nii
        |--  sample21_normal_img.nii.gz
        |--  sample21_normal_mask.nii
                .
                .
                .   

3. Upstream

📋 Available List

• Up_SMART_Net
• Up_SMART_Net_Dual_CLS_SEG
• Up_SMART_Net_Dual_CLS_REC
• Up_SMART_Net_Dual_SEG_REC
• Up_SMART_Net_Single_CLS
• Up_SMART_Net_Single_SEG
• Up_SMART_Net_Single_REC

+ train: We conducted upstream training with three multi-task including classificatiom, segmentation and reconstruction.

python train.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--model-name 'Up_SMART_Net' \
--batch-size 10 \
--epochs 1000 \
--num-workers 4 \
--pin-mem \
--training-stream 'Upstream' \
--multi-gpu-mode 'DataParallel' \
--cuda-visible-devices '2, 3' \
--gradual-unfreeze 'True' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test'

+ test (slice-wise for slice-level): We conducted upstream training with three multi-task including classificatiom, segmentation and reconstruction.

python test.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner "True" \
--model-name 'Up_SMART_Net' \
--num-workers 4 \
--pin-mem \
--training-stream 'Upstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test/epoch_0_checkpoint.pth'

+ test (stacking slice for volume-level): We conducted upstream training with three multi-task including classificatiom, segmentation and reconstruction.

python test.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner "False" \
--model-name 'Up_SMART_Net' \
--num-workers 4 \
--pin-mem \
--training-stream 'Upstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test/epoch_0_checkpoint.pth'

4. Downstream

📋 Available List

• Down_SMART_Net_CLS
• Down_SMART_Net_SEG

- SMART-Net-CLS

+ train: We conducted downstream training using multi-task representation.

python train.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--model-name 'Down_SMART_Net_CLS' \
--batch-size 2 \
--epochs 1000 \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'DataParallel' \
--cuda-visible-devices '2, 3' \
--gradual-unfreeze 'True' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_cls_test' \
--from-pretrained '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/[UpTASK]ResNet50_ImageNet.pth' \
--load-weight-type 'encoder'

+ test: We conducted upstream training with three multi-task including classificatiom, segmentation and reconstruction.

python test.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner 'False' \
--model-name 'Down_SMART_Net_CLS' \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_cls_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_cls_test/epoch_0_checkpoint.pth'

- SMART-Net-SEG

+ train: We conducted downstream training using multi-task representation.

python train.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--model-name 'Down_SMART_Net_SEG' \
--batch-size 2 \
--epochs 1000 \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'DataParallel' \
--cuda-visible-devices '2, 3' \
--gradual-unfreeze 'True' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_seg_test' \
--from-pretrained '/workspace/sunggu/1.Hemorrhage/SMART-Net/up_test/epoch_0_checkpoint.pth' \
--load-weight-type 'encoder'

+ test: We conducted upstream training with three multi-task including classificatiom, segmentation and reconstruction.

python test.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner 'False' \
--model-name 'Down_SMART_Net_SEG' \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_seg_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_seg_test/epoch_0_checkpoint.pth'

SMART-Net Inference

📋 Available List

• Up_SMART_Net
• Down_SMART_Net_CLS
• Down_SMART_Net_SEG

- Up-SMART-Net

+ inference: We conducted downstream training using multi-task representation.

python inference.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner "True" \
--model-name 'Up_SMART_Net' \
--num-workers 4 \
--pin-mem \
--training-stream 'Upstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/up_test/epoch_0_checkpoint.pth'

- Down-SMART-Net-CLS

+ inference: We conducted downstream training using multi-task representation.

python inference.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner "False" \
--model-name 'Down_SMART_Net_CLS' \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_cls_test' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_cls_test/epoch_0_checkpoint.pth'

- Down-SMART-Net-SEG

+ inference: We conducted downstream training using multi-task representation.

python inference.py \
--data-folder-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/datasets/samples' \
--test-dataset-name 'Custom' \
--slice-wise-manner "False" \
--model-name 'Down_SMART_Net_SEG' \
--num-workers 4 \
--pin-mem \
--training-stream 'Downstream' \
--multi-gpu-mode 'Single' \
--cuda-visible-devices '2' \
--print-freq 1 \
--output-dir '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_seg_test/pred_nii' \
--resume '/workspace/sunggu/1.Hemorrhage/SMART-Net/checkpoints/down_seg_test/epoch_0_checkpoint.pth'

Upstream visualize

1. Activation map

⏳ It's scheduled to be uploaded soon.

2. t-SNE

⏳ It's scheduled to be uploaded soon.

Excuse

For personal information security reasons of medical data in Korea, our data cannot be disclosed.

📝 Citation

If you use this code for your research, please cite our papers:

⏳ It's scheduled to be uploaded soon.

🤝 Acknowledgement

We build SMART-Net framework by referring to the released code at qubvel/segmentation_models.pytorch and Project-MONAI/MONAI. This is a patent-pending technology.

🛡️ License

Project is distributed under MIT License