ModDrop++

The repository is the official PyTorch implementation of the paper "ModDrop++: A Dynamic Filter Network with Intra-subject Co-training for Multiple Sclerosis Lesion Segmentation with Missing Modalities". [paper]

Modality Dropout (ModDrop) has been widely used as an effective training scheme to train a unified model that can be self-adaptive to different missing conditions. However, the classic ModDrop suffers from two limitations: (1) regardless of different missing conditions, it always forces the network to learn a single set of parameters and thus may limit the expressiveness of the network and (2) ModDrop does not leverage the intra-subject relation between full- and missing-modality data. To address these two limitations, the proposed ModDrop++ incoportates (1) a plug-and-play dynamic head and (2) an intra-subject co-training strategy to upgrade the ModDrop. ModDrop++ has been developed and implemented based on the 2.5D Tiramisu model, which achieved the state-of-the-art performance for MS lesion segmentation on the ISBI 2015 challenge.

The trained models for UMCL and ISBI datasets are available here.

If you find our code/paper helpful for your research, please consider citing our work:

@inproceedings{liu2022moddrop++,
  title={ModDrop++: A Dynamic Filter Network with Intra-subject Co-training for Multiple Sclerosis Lesion Segmentation with Missing Modalities},
  author={Liu, Han and Fan, Yubo and Li, Hao and Wang, Jiacheng and Hu, Dewei and Cui, Can and Lee, Ho Hin and Zhang, Huahong and Oguz, Ipek},
  booktitle={Medical Image Computing and Computer Assisted Intervention--MICCAI 2022: 25th International Conference, Singapore, September 18--22, 2022, Proceedings, Part V},
  pages={444--453},
  year={2022},
  organization={Springer}
}

If you have any questions, feel free to contact [email protected] or open an Issue in this repo.

Prerequisites

• NVIDIA GPU + CUDA + cuDNN

Installation

We suggest installing the dependencies using Anaconda

• create the environment and activate (replace DL with your environment name)

conda create --name DL python=3.8

• Install PyTorch with the official guide from http://pytorch.org (we used the CUDA version 10.0), and then install other dependencies:

conda activate DL
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
conda install nibabel statsmodels visdom jsonpatch dominate scikit-image -c conda-forge -c anaconda

Datasets

You can put your images in any folder, but to run this model, we assume the format of file names in the pattern like {PREFIX}_{PATIENT-ID}_{TIMEPOINT-ID}_{MODALITY(MASK)}.{SUFFIX}, e.g. training_01_01_flair.nii, training_03_05_mask1.nii :

PREFIX: can be any string
PATIENT-ID: number id of the patient
TIMEPOINT-ID: number id of the timepoint
MODALITY, MASK: t1, flair, t2, pd, mask1, mask2, etc
SUFFIX: nii, nii.gz, etc

You need to specify the name of MODALITIES, MASKS and SUFFIX in the configuration.py under the root folder.

Data Conversion

This data conversion is to convert the different kinds of datasets into the same structure. To illustrate the process of training and validation, we use the ISBI-2015 challenge data as an example. The steps are as follows:

1. Copy all of the preprocessed training data of ISBI challenge dataset into the target dataset folder. The folder is ~/Documents/Datasets/sample_dataset

2. Modify the configuration file (configuration.py). The positions that need attention are marked with TODO.

   • The dataset folder, e.g. PATH_DATASET = os.path.join(os.path.expanduser('~'), 'Documents', 'Datasets', 'sample_dataset').
   • The modalities available with this dataset, e.g. MODALITIES = ['t1', 'flair', 't2', 'pd'].
   • The delineations available with this dataset, e.g. MASKS = ['mask1', 'mask2'].
   • The suffix of the files, usually 'nii' or 'nii.gz'
   • (This is not necessary for now) The axis corresponding to axial, sagittal and coronal, respectively. For the ISBI dataset, it is [2, 0, 1].

3. Rename all the files to comply with the pattern as mentioned in the Dataset section. For example, since we use 't1' in the MODALITIES, simply rename 'training01_01_mprage_pp.nii' to 'training_01_01_t1.nii'. The following commands that might be useful:

   rename 's/training/training_/g' *; rename 's/_pp.nii/.nii/g' *; rename 's/mprage/t1/g' *; 

4. Run the data_conversion file. This function will move your files under the sample_dataset folder into its subfolder ./raw and generate two JSON files (ids.json, properties.json).

   The ids.json contains the new data paths based on PATIENT-ID, TIMEPOINT-ID, MODALITY or MASK. It is needed for cross-validation to avoid splitting the scan from the same patient into both Train and Val (or Test) folder. An example is

   "1": {
    "1": {
      "modalities": {
        "t1": "~/Documents/Datasets/sample_dataset/raw/training_01_01_t1.nii",
        "flair": "~/Documents/Datasets/sample_dataset/raw/training_01_01_flair.nii",
        "t2": "~/Documents/Datasets/sample_dataset/raw/training_01_01_t2.nii",
        "pd": "~/Documents/Datasets/sample_dataset/raw/training_01_01_pd.nii"
      },
      "mask": {
        "mask1": "~/Documents/Datasets/sample_dataset/raw/training_01_01_mask1.nii",
        "mask2": "~/Documents/Datasets/sample_dataset/raw/training_01_01_mask2.nii"
      }
    },
    "2": {
    ...
    }
    ...
   }
   

   The properties.json contains the peak of each modality using kernel density estimation. It is saved in the JSON file so that we don't need to calculate it repeatedly during the training process. An example is

   "406576b1b92f6740b0e20a29016952ae1fa6c4cf": {
   "path": "~/Documents/Datasets/sample_dataset/raw/training_01_01_t1.nii",
   "peak": 178855.95956321745
   }

Training and Validation

Before you run and training and validation, you can decide what kind of cross-validation strategy you want to use. At default, we use 5-fold cross-validation. We implemented three strategies, which can be set using test_mode option:

• 'val': validation only, no test. For each fold, 4/5 of data for training and 1/5 of data for validation. The models with the best dice score will be saved as 'latest' models. 5 'latest' models will be preserved. The training will stop after 160 epoch of no improvement or the fixed number of epochs as provided in the training options, whichever comes first. Since all the ISBI data we placed into sample_dataset are from the training dataset (which means the test data will be provided separately), we use this mode.
• 'test': test only, no validation. The fold can be set using test_index option. Only one model will be generated. The program will not automatically save the 'latest' model because the model performance is determined using the validation set. The program will run fixed epochs and stop.
• 'val_test': do both validation and test. The test fold is set to the last fold at default (1/5 of data) and will not be seen by the model during the training/validation process. For the remaining 4/5 of data, training takes 3/5 and validation takes 1/5. In such a way, 4 'latest' model will be saved and they will be finally tested with the hold test set.

The visdom is needed to run before the training function if you want to visualize the results

conda activate DL
python -m visdom.server

An example to run the ModDrop++ training is

conda activate DL
python train.py --loss_to_use focal_ssim --input_nc 3 --trainSize 128 --test_mode val_test --name experiment_name --eval_val --batch_size 16 

where:

• loss_to_use: you can choose from ['focal', 'l2', 'dice', 'ssim']. You can also use combinations of them, e.g. 'focal_l2'. If you use more than one loss functions, you can set the weight of each loss by setting lambda_L2, lambda_focal and lambda_dice.
• input_nc: the number of slices in a stack. As mentioned in the paper, using 3 achieved the best results.
• trainSize: for ISBI dataset, the size is 181x217x181. We set the trainSize to be 128 so that the program will crop the slice during the training process. If you set it to 256, the program will pad the slices.
• name: the experiment name. The checkpoint of this name will be automatically created under the ../Checkpoint folder. Also, in the test phase, the mask prediction has the same suffix as this value. *eval_val: use the eval mode for validation during the training process.

If you got "RuntimeError: unable to open shared memory object", use PyCharm to run the code instead of using terminal

Testing

In the testing phase, you can create a new folder for the test dataset. For example, if you want to use the test set of ISBI challenge, create a new subfolder ./challenge under sample_dataset (parallel to train, val, test). The naming of files does not need to follow the strict pattern described in Dataset and Data Conversion, but the files should end with {MODALITY(MASK)}.{SUFFIX}. The reference mask files are not needed. If the reference mask files are provided, the segmentation metrics will be output.

Remember to switch to Testing mode in 'models/ms_model.py' file. This needs to be done manually for now and will be cleaned up later.

Example:

conda activate DL
python test.py --load_str val0test4 --input_nc 3 --testSize 512 --name experiment_name --epoch 300 --phase test --eval

where:

• load_str: in the training phase, the lastest models are saved due to cross-validation. This string describes the models (folds) you want to load. With it is set to 'val0,val1,val2,val3,val4', latest models from all the 5 folds are loaded (you can check the model names by exploring the checkpoint folder).
• testSize: the default value is 256. The slices will be padded to this size in the inference stage. If you have any dimension larger than 256, simply change it to a number that larger than all the dimension sizes.
• epoch: the default value is latest. You can set it to a specific number, but it is not recommended.
• phase: the new subfolder of the test set.

You can find the options files under the option folder to get more flexibility of running the experiments.