ai-med / daft

Dynamic Affine Feature Map Transform

Combining 3D Image and Tabular Data via the Dynamic Affine Feature Map Transform

This repository contains the code to the paper "DAFT: A Universal Module to Interweave Tabular Data and 3D Images in CNNs"

@article{Wolf2022-daft,
  title = {{DAFT: A Universal Module to Interweave Tabular Data and 3D Images in CNNs}},
  author = {Wolf, Tom Nuno and P{\"{o}}lsterl, Sebastian and Wachinger, Christian},
  journal = {NeuroImage},
  pages = {119505},
  year = {2022},
  issn = {1053-8119},
  doi = {10.1016/j.neuroimage.2022.119505},
  url = {https://www.sciencedirect.com/science/article/pii/S1053811922006218},
}

and the paper "Combining 3D Image and Tabular Data via the Dynamic Affine Feature Map Transform"

@inproceedings(Poelsterl2021-daft,
  title     = {{Combining 3D Image and Tabular Data via the Dynamic Affine Feature Map Transform}},
  author    = {P{\"{o}}lsterl, Sebastian and Wolf, Tom Nuno and Wachinger, Christian},
  booktitle = {International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  pages     = {688--698},
  year      = {2021},
  url       = {https://arxiv.org/abs/2107.05990},
  doi       = {10.1007/978-3-030-87240-3_66},
}

If you are using this code, please cite the papers above.

Installation

1. Use conda to create an environment called daft with all dependencies:

conda env create -n daft --file requirements.yaml

2. (Optional) If you want to run the ablation study, you need to setup a Ray cluster to execute the experiments in parallel. Please refer to the Ray documentation for details.

Data

We used data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Since we are not allowed to share our data, you would need to process the data yourself. Data for training, validation, and testing should be stored in separate HDF5 files, using the following hierarchical format:

1. First level: A unique identifier, e.g. image ID.
2. The second level always has the following entries:
   1. A group named Left-Hippocampus, which itself has the dataset named vol_with_bg as child: The cropped ROI around the left hippocampus of size 64x64x64.
   2. A dataset named tabular of size 14: The tabular non-image data.
   3. A scalar attribute RID with the patient ID.
   4. A string attribute VISCODE with ADNI's visit code.
   5. Additional attributes, depending on the task.
      1. For classification, a string attribute DX containing the diagnosis: CN, MCI, or Dementia.
      2. For time-to-dementia analysis. A string attribute event indicating whether conversion to dementia was observed (yes or no), and a scalar attribute time with the time to dementia onset or the time of censoring.

One entry in the resulting HDF5 file should have the following structure:

/1010012                 Group
    Attribute: RID scalar
        Type:      native long
        Data:  1234
    Attribute: VISCODE scalar
        Type:      variable-length null-terminated UTF-8 string
        Data:  "bl"
    Attribute: DX scalar
        Type:      variable-length null-terminated UTF-8 string
        Data:  "CN"
    Attribute: event scalar
        Type:      variable-length null-terminated UTF-8 string
        Data:  "no"
    Attribute: time scalar
        Type:      native double
        Data:  123
/1010012/Left-Hippocampus Group
/1010012/Left-Hippocampus/vol_with_bg Dataset {64, 64, 64}
/1010012/tabular         Dataset {14}

Finally, the HDF5 file should also contain the following meta-information in a separate group named stats:

/stats/tabular           Group
/stats/tabular/columns   Dataset {14}
/stats/tabular/mean      Dataset {14}
/stats/tabular/stddev    Dataset {14}

They are the names of the features in the tabular data, their mean, and standard deviation.

Usage

Training

To train DAFT, or any of the baseline models, execute the train.py script to begin training. The essential command line arguments are:

• --task: The type of loss to optimize. Can be clf for classification, and surv for time-to-dementia analysis.
• --train_data: Path to HDF5 file containing training data.
• --val_data: Path to HDF5 file containing validation data.
• --test_data: Path to HDF5 file containing test data.

Model checkpoints will be written to the experiments_clf or experiments_surv folder, depending on the value of --task.

For a full list of command line arguments, execute:

python train.py --help

Ablation Study

Performing the ablation study is computationally quite expensive, and requires a Ray cluster to be setup to execute experiments in parallel. Please refer to the Ray documentation on how to do that. We expect that ray-cluster.yaml refers to a valid Ray cluster config.

There are 2 scripts that you can execute:

• ablation_adni_classification.py: Ablation study for classification experiments. Will create results_adni_classification_ablation_split-*.csv files on the Ray head node.
• ablation_adni_survival.py: Ablation study for time-to-dementia experiments. Will create results_adni_survival_ablation_split-*.csv files on the Ray head node.

These are the steps you need to execute:

1. Start the Ray cluster:

ray up ray-cluster.yaml

2. Execute the ablation_adni_classification.py (or ablation_adni_survival.py):

ray exec ray-cluster.yaml --tmux \
  '/usr/bin/python /path/to/workspace/ablation_adni_classification.py --data_dir=/path/to/data/ --ray_address=<hostname>:<port>'

3. Monitor the progress:

ray attach ray-cluster.yaml --tmux

4. Inspect the final results for the first fold:

cat results_adni_classification_ablation_split-0.csv

daft's Issues

Issue Regarding Data Loading

Hello, author.
I have processed my data into the HDF5 data structure as described in your README. However, when running the _def load(self, filename, dataset_name, roi="Left-Hippocampus") function, I always encounter the error KeyError: "Unable to open object (object 'Left-Hippocampus' doesn't exist)" when reaching the line _meta = self.get_meta_data(hf["stats"][roi][dataset_name]). To address this, I added a sub-group named "Left-Hippocampus" in the "stats" group and set its "vol_with_bg" value to (64, 64, 64). However, this resulted in another error: AttributeError: 'Dataset' object has no attribute 'items'.
So, I would like to seek your advice again regarding the data structure. Is there a need to add something specific in the "stats" group?

A tiny question about the data processing

Dear authors,

Thanks for the great work!

I am writing to ask you about the tabular data processing strategy in your paper. It is very interesting to add a binary variable indicating the missingness of a tabular value. I kindly wonder besides the binary variable, how did you fill in the value of the missing tabular variables? Should it be somewhat 0 or imputed by interpolation/KNN? I am asking since I suppose we might need to feed them into the network. 

Issue regarding data loading

Hi
I have processed my data into the HDF5 data structure as described in your README. However, when running the _def load(self, filename, dataset_name, roi="Left-Hippocampus") function, I always encounter the error KeyError: "Unable to open object (object 'Left-Hippocampus' doesn't exist)" when reaching the line _meta = self.get_meta_data(hf["stats"][roi][dataset_name]).
I appreciate your response for solving this issue.