Deep learning based registration using spatial gradients and noisy segmentation labels (Learn2Reg Task 3: CT Abdominal)
[Edit 12/21] : Models are now available on Zenodo for both hippocampus and abdominal registration
This repository contains a Pytorch implementation of Deep learning based registration using spatial gradients and noisy segmentation labels. It corresponds to the 2nd ranked for Task 3 (CT Abdominal) and 2nd overall method for the Learn2Reg Challenge 2020 : https://learn2reg.grand-challenge.org/.
The presentation of our method is available on the Learn2Reg website : https://cloud.imi.uni-luebeck.de/s/FJ3szqokbZRfjzj. Presentation from other participants to the workshop are also available : https://learn2reg.grand-challenge.org/Workshop/.
You can also consult the repository for the Task 4 : https://github.com/TheoEst/hippocampus_registration. Implementation is made with Pytorch.
In order to use this repository, you only need to download the Learn2Reg Task 3 Data : https://learn2reg.grand-challenge.org/Datasets/ and add it on the ./data/ folder.
If you want to use the supplementary data, you need to download the following datasets :
- Medical Segementation Decathlon (http://medicaldecathlon.com/). We use 5 cohorts from the MSD dataset : Liver, Pancreas, Colon, Hepatic Vessels and Spleen.
- KITS 19 (https://github.com/neheller/kits19).
- TCIA Pancreas (https://wiki.cancerimagingarchive.net/display/Public/Pancreas-CT and https://zenodo.org/record/1169361#.X4Vr2efgqUk).
For each of the supplementary dataset, you need to run the two code ./preprocessing/resample_cohort_name.py and ./preprocessing/ants_warped_cohort_name.py . Theses preprocessing steps will perform resampling to 2 mm voxel and linear registration with AntsPy (https://github.com/ANTsX/ANTsPy).
Our method is based on the article Deep Learning-Based Concurrent Brain Registration and Tumor Segmentation, Estienne T., Lerousseau M. et al., 2020 (https://www.frontiersin.org/articles/10.3389/fncom.2020.00017/full).
In this work we proposed a deep learning based registration using 3D Unet as backbone with 3 losses :
- Reconstruction loss ( Mean Square Error or Local Cross Correlation)
- Segmentation loss ( Dice Loss between deformed segmentation and ground truth segmentation)
- Regularisation loss (To force smoothness)
In the proposed architecture, the moving and fixed image are passed independently through the encoder, and then merged with subtraction operation.
Models can be download on Zenodo on the following link : https://zenodo.org/record/5762347#.Ya5u4edCewA
5 pretrained models are available on the ./models folder :
- Baseline model
- Baseline model with symmetric training
- Baseline model with pretraining model (unsupervised pretraining)
- Baseline model with pretraining model (supervised pretraining with noisy segmentations)
- Baseline model with pretraining model (supervised pretraining with noisy segmentations), trained with both training and validation dataset (used for the test submission)
To recreate this models, launch the following commands :
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Baseline --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --deformed-mask-loss --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --keep-all-label
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Baseline+symmetric --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --deformed-mask-loss --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --keep-all-label --symmetric-training
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Baseline+symmetric+pretrain --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --deformed-mask-loss --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --keep-all-label --symmetric-training --model-abspath ./abdominal_registration/save/models/Pretrain_unsupervised.pth.tar
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Baseline+symmetric+pretrain_noisy_labels --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --deformed-mask-loss --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --keep-all-label --symmetric-training --model-abspath ./abdominal_registration/save/models/Pretrain_supervised.pth.tar
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Test_submission--lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --deformed-mask-loss --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --keep-all-label --symmetric-training --merge-train-val --model-abspath ./abdominal_registration/save/models/Pretrain_supervised.pth.tar
Two pretraining models are given : unsupervised pretraining and supervised with noisy labels. The pretraining consist in training with the supplementary datasets and used the noisy labels (called pseudo labels in the code).
To recreate this models, launch the following commands :
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Pretrain_unsupervised --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --symmetric-training --cohorts learn2reg_task3 kits liver tcia_pancreas spleen colon pancreas hepatic --val-cohorts learn2reg_task3 --pseudo-labels
python3 -m abdominal_registration.main --crop-size 128 128 128 --val-crop-size 192 160 192 --zeros-init --workers=4 --batch-size=2 --val-batch-size 1 --epochs=300 --session-name=Pretrain_supervised --lr=1e-4 --instance-norm --regu-deformable-loss-weight=1e-2 --mse-loss-weight=1 --random-crop --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --symmetric-training --cohorts learn2reg_task3 kits liver tcia_pancreas spleen colon pancreas hepatic --val-cohorts learn2reg_task3 --pseudo-labels --deformed-mask-loss
To predict, use the predict_reg.py file.
Options :
--val Do the inference for the validation dataset
--train Do the inference for the train dataset
--test Replace the validation dataset by test set. (--val is necessary)
--save-submission Save the submission in the format for the Learn2Reg challenge
--save-deformed-img Save the deformed image and deformed mask in numpy format
--save-grid Save the grid in numpy format
Examples :
python3 -m abdominal_registration.predict_reg --crop-size 256 160 192 --batch-size=1 --val-batch-size 1 --instance-norm --classic-vnet --channel-multiplication 8 --deep-supervision --multi-windows --cohorts learn2reg_task3 --cohorts learn2reg_task3 --val --train --save-submission --model-abspath ./abdominal_registration/save/models/Baseline+symmetric+pretrain.pth.tar
We developped a segmentation network based on a 3D Unet in order to segment 11 abdominal organs. The goal is to predict the segmentations for the supplementary data and used its for the pretraining. Our model segment the 11 following organs : Spleen, Right & Left Kidney, Liver, Stomach, Pancreas, Gall Bladder, Aorta, Inferior Vena Cava, Portal & Splenic Vein, Esophagus. The segmentation network was trained with the following cohorts : Learn2Reg Task3, TCIA Pancreas, KITS 19 and MSD (Liver, Spleen and Pancreas). We used a modified dice loss function to train our network, such that we backpropagate only the loss for the organs contained in the cohort.
Two scripts are available to train the segmentation model : main_seg.py and predict_seg.py. To recreate our experiments, launch the following commands :
python3 -m abdominal_registration.main_seg --batch-size=3 --crop-size 144 144 144 --cohorts learn2reg_task3 liver pancreas spleen tcia_pancreas kits --random-crop --lr=1e-4 --instance-norm --session-name Train_seg_multi_cohorts --val-crop 192 160 192 --classic-vnet --data-augmentation --epochs=400 --val-cohorts tcia_pancreas learn2reg_task3
python3 -m abdominal_registration.predict_seg --batch-size=1 --crop-size 256 160 192 --cohorts learn2reg_task3 liver pancreas spleen tcia_pancreas kits hepatic colon --instance-norm --classic-vnet --train --val --model-abspath ./abdominal_registration/save/models/Train_seg.pth.tar
To transform the predicted data into a compressed file, just use the create_submission.py file. For instance python3 ./submission/create_submission.py ./save/submission/Baseline+symmetric+pretrain . You will obtain a folder called Baseline+symmetric+pretrain_compressed and a zip file Baseline+symmetric+pretrain_submission which you can submit.
To evaluate the performance, you need just to run the apply_evaluation.py file. For instance python3 ./submission/apply_evaluation.py Baseline+symmetric+pretrain_compressed will generate a csv file in the ./save/evaluation/ folder with all the metrics for each pairs (Dice, Dice30, Hausdorff and standard deviation of Jacobian).
Results on the validation set
| Method | Dice | Dice 30 | Hausdorff Distance | Jacobian |
|---|---|---|---|---|
| Unregistered | 0.23 | 0.01 | 46.1 | |
| Baseline | 0.384 | 0.346 | 45.2 | 1.70 |
| Baseline + sym. | 0.400 | 0.358 | 45.7 | 1.67 |
| Baseline + sym. + pretrain | 0.522 | 0.501 | 42.3 | 0.327 |
| Baseline + sym. + pretrain + noisy labels | 0.625 | 0.580 | 39.2 | 1.77 |
| Test | 0.64 | 0.40 | 37.1 | 1.53 |
Example of the results on the validation set :
This package was developed by Théo Estienne12
1 Université Paris-Saclay, CentraleSupélec, Mathématiques et Informatique pour la Complexité et les Systèmes, 91190, Gif-sur-Yvette, France.
2 Université Paris-Saclay, Institut Gustave Roussy, Inserm, Radiothérapie Moléculaire et Innovation Thérapeutique, 94800, Villejuif, France.