Source-Free Domain Adaptation

New: Please check out SFDA, the repository of our Source-Free Domain Adaptation method which will be updated.

Mathilde Bateson, Hoel Kervadec, Jose Dolz, Hervé Lombaert, Ismail Ben Ayed @ETS Montréal

Code of our submission at MICCAI 2020 and its ongoing journal extension. Video of the MICCAI talk is available: https://www.youtube.com/watch?v=ALYaa5xrxbQ&ab_channel=MB

• MICCAI 2020 Proceedings
• arXiv preprint

Please cite our paper if you find it useful for your research.

@inproceedings{BatesonSFDA,
	Author = {Bateson, Mathilde and Kervadec, Hoel and Dolz, Jose and Lombaert, Herv{\'e} and Ben Ayed, Ismail},
	Booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2020},
	Pages = {490--499},
	Publisher = {Springer International Publishing},
	Title = {Source-Relaxed Domain Adaptation for Image Segmentation},
	Year = {2020}
    Address = {Cham}}

Requirements

Non-exhaustive list:

• python3.6+
• Pytorch 1.0
• nibabel
• Scipy
• NumPy
• Matplotlib
• Scikit-image
• zsh
• tqdm
• pandas
• scikit-image

Data scheme

datasets

For instance

data
    prostate_source/
	    train/
		IMG/
		    Case10_0.png
		    ...
		GT/
		    Case10_0.png
		    ...
		...
	    val/
		IMG/
		    Case11_0.png
		    ...
		GT/
		    Case11_0.png
		    ...
		...
    prostate_target/
	    train/
		IMG/
		    Case10_0.png
		    ...
		GT/
		    Case10_0.png
		    ...
		...
	    val/
		IMG/
		    Case11_0.png
		    ...
		GT/
		    Case11_0.png
		    ...
		...

The network takes png or nii files as an input. The gt folder contains gray-scale images of the ground-truth, where the gray-scale level is the number of the class (0,1,...K).

Class-ratio (sizes) prior

The class-ratio prior is estimated from anatomical knowledge for each application. In our implementation, is it estimated for each slice in the target domain training and validation sets. It estimated once, before the start of the adaptation phase, and saved in a csv file.

Scheme

sizes/
    prostate.csv
    whs.csv
    ivd.csv

The size csv file should be organized as follows:

val_ids	dumbpredwtags
Case00_0.nii	[Estimated_Size_class0, Estimated_Size_class1, ..., Estimated_Size_classk]

Sample from sizes/prostate.csv :

val_ids	val_gt_size	dumbpredwtags
Case00_0.nii	[147398.0, 827.0]	[140225, 6905]
Case00_1.nii	[147080.0, 1145.0]	[140225, 6905]
Case00_14.nii	[148225.0, 0.0]	[148225, 0]

NB 1 : there should be no overlap between names of the slices in the training and validation sets (Case00_0.nii,...).

NB 2: in our implementation, the csv file contains the sizes priors in pixels, and the KL Divergence loss divides the size in pixels by (w*h) the height and weight of the slice, to obtain the class-ratio prior.

NB 3: Estimated_Size_class0 + Estimated_Size_class1 + ... + Estimated_Size_classk = w*h

NB 4: the true val_gt_size is unknown, so it is not directly used in our proposed SFDA. However, in our framework an image-level annotation is available for the target training dataset: the "Tag" of each class k, indicating the presence or absence of class k in the slice. Therefore, Estimated_Size_classk=0 if val_gt_size_k = 0 and Estimated_Size_classk>0 if val_gt_size_k > 0

NB 5: To have an idea of the capacity of the SFDA model in the ideal case where the ground truth class-ratio prior is known, it is useful to run the upper bound model SFDA_TrueSize choosing the column "val_gt_size" instead of "dumbpredwtags". This can be changed in the makefile :

results/sa/SFDA_TrueSize: OPT = --target_losses="[('EntKLProp', {'lamb_se':1,'lamb_consprior':1,'ivd':True,'weights_se':[0.1,0.9],'idc_c': [1],'curi':True,'power': 1},'PredictionBounds', \
      {'margin':0,'dir':'high','idc':[0,1],'predcol':'val_gt_size','power': 1, 'mode':'percentage','sizefile':'sizes/prostate.csv'},'norm_soft_size',1)]" \
           --val_target_folders="$(TT_DATA)"  --l_rate 0.000001 --n_epoch 100 --lr_decay 0.9 --batch_size 10 --target_folders="$(TT_DATA)" --model_weights="$(M_WEIGHTS_ul)" \

NB 6 : If you change the name of the columns (val_ids, dumbpredwtags) in the size file, you should change them in the bounds.py file as well as in the prostate.make makefile.

results

results/
    prostate/
        fs/
            best_epoch_3d/
                val/
                    Case11_0.png
                    ...
            iter000/
                val/
            ...
        sfda/
            ...
        params.txt # saves all the argparse parameters of the model 
	best_3d.pkl # best model saved
	last.pkl # last epoch
        IMG_target_metrics.csv # metrics over time, csv
        3dbestepoch.txt # number and 3D Dice of the best epoch 
        ...
    whs/
        ...
archives/
    $(REPO)-$(DATE)-$(HASH)-$(HOSTNAME)-sfda.tar.gz
    $(REPO)-$(DATE)-$(HASH)-$(HOSTNAME)-prostate.tar.gz

Interesting bits

The losses are defined in the losses.py file. See below is the general structure of the target_losses parameter. The lambda weight of the loss is the last parameter. parser.add_argument("--target_losses", type=str, required=True, help="List of (loss_name, loss_params, bounds_name, bounds_params, fn, weight)")

Example for the supervised Cross-Entropy loss:

target_losses="[('CrossEntropy', {'idc': [0,1], 'weights':[1,1]}, None, None, None, 1)]"

Example for our SFDA loss on prostate:

target_losses="[('EntKLProp', {'lamb_se':1, 'lamb_consprior':1,'ivd':True,'weights_se':[0.1,0.9],
'idc_c': [1],'curi':True,'power': 1},'PredictionBounds', \
{'margin':0,'dir':'high','idc'[0,1],'predcol':'dumbpredwtags','power':1,
'mode':'percentage','sep':';','sizefile':'sizes/prostate.csv'},'norm_soft_size',1)]" 

Running our main experiment

Once you have downladed the data and organized it such as in the scheme above:

Change the path to match the base folder containing the target dataset in the makefiles : T_FOLD = "your/path/to/the/target/dataset"

Run the main experiments as follows:

make -f prostate.make 
make -f ivd.make 
make -f whs.make 

This will first run the SFDA model, which will be saved in results/sfda. It will be initialized with the networks weights obtained by training on the source only, which we made available. Eventually, you can re-run the source training model, which will be saved in results/prostate/cesource.

Cool tricks

Remove all assertions from the code to speed up. Usually done after making sure it does not crash for one complete epoch:

make -f prostate.make <anything really> CFLAGS=-O

Use a specific python executable:

make -f prostate.make <super target> CC=/path/to/the/executable

Train for only 5 epochs, with a dummy network, and only 10 images per data loader. Useful for debugging:

make -f prostate.make <anything really> NET=Dimwit EPC=5 DEBUG=--debug

Rebuild everything even if already exist:

make -f prostate.make <a> -B

Only print the commands that will be run (useful to check recipes are properly defined):

make -f prostate.make <a> -n

Tip for implementing on a new dataset

This is list of parameters and hyperparameters which have influence the model's quality:

• the class-ratio anatomical prior (defined in --target_losses)
• the lambda weight for the KL loss (defined in --target_losses)
• the class weights for the KL loss (defined in --target_losses)
• the class weights for the entropy minimization loss (defined in --target_losses)
• the learning rate, and the learning rate decay

Regarding the class-ratio anatomical prior: Using the ground truth slice-based class-ratio prior (val_gt_size in the size file) is a good idea to get an idea of the capability of the SFDA on your application (it is an upper bound for our SFDA). Our main SFDA method relies on obtaining an estimation of the class-ratio for each structure of interest. Empirically, we found that an over estimation of the size of a structure yielded better quantitative and qualitative results than an under estimation^{1}. Therefore, on a new application, obtaining a coarse upper bound of an anatomical structure's pixel size should be a good starting point.

Related Implementation and Dataset

• Mathilde Bateson, Hoel Kervadec, Jose Dolz, Hervé Lombaert, Ismail Ben Ayed. Constrained Domain Adaptation for Image Segmentation. In IEEE Transactions on Medical Imaging, 2021. [paper] [implementation]
• Hoel Kervadec, Jose Dolz, Meng Tang, Eric Granger, Yuri Boykov, Ismail Ben Ayed. Constrained-CNN losses for weakly supervised segmentation. In Medical Image Analysis, 2019. [paper] [code]
• Prostate Dataset and details: https://raw.githubusercontent.com/liuquande/SAML/. The SA site dataset was used a target domain, the SB site was used as source domain. For both datasets, we use 20 scans for training, and the remaining 10 scans for validation.
• Heart Dataset and details: We used the preprocessed dataset from Dou et al. : https://github.com/carrenD/Medical-Cross-Modality-Domain-Adaptation. The data is in tfs records, it should be transformed to nii or png before running the makefile.
• Spine Dataset and details: https://ivdm3seg.weebly.com/ . From the original coronal view, we transposed the slices to transverse view in our experiments. We set the water modality (Wat) as the source and the in-phase (IP) modality as the target domain. From this dataset, 13 scans are used for training, and the remaining 3 scans for validation.

Note

The model and code are available for non-commercial research purposes only.

{1} This could be traced to the well-known under segmentation bias of the entropy minimization alone (see our paper), which seems better compensated when the KL matching has an over segmentation bias. Note that the common DICE metric also favors oversegmentation over undersegmentation. ( See https://openreview.net/pdf?id=76X9Mthzv4X for example)