This dataset contains abdominal MRI series for 105 patients, who underwent MRI examination at Duke University Hospital. For each patient there is exactly 1examination, and for each examination there are tens of series. In each series,there are also tens of MRI slices (stored as dicom format file). An illustration of the hierarchy of the dataset is shown in the figure below.

We provide all the 30 series types in the following table. In each cell the format is series type(numeric label).

This dataset has been deidentified before publication, which means it contains no PHI.

###Dependencies

• tensorflow
• opencv-python
• pydicom The version of the libraries we tested are: tensorflow 1.15.0, opencv 4.2.0.32 and pydicom 1.4.2. You can install those by:

pip install tensorflow-gpu==1.15 
pip install opencv-python==4.2.0.32
pip install pydicom==1.4.2

A pretrained model (checkpoint) is in the 0621 directory. The test image is an anonymized abdominal liver MRI:

Then you can run inception_2d_test.py and the script will output:

Predicted Series Type: arterial
Probability: 0.718601822853

Since the whole dataset is about 20GB, loading it to the memory is infeasible for a typical PC. We train the model using tensorflow Estimator and convert the images and labels into tfrecord format, so that during training when a batch is needed it can be loaded efficiently. Note that for demonstration purpose we only adopt a 2D model for simplicity.

First, you need to change to directories in the construct_tfrecord_2d.py. In line 83 change dataset_folder to the directory where you put the dataset, and in line 85 change the output_folder to the directory where you would like the generated file to store. After these two values are set, you can just run the script.

Run the inception_2d_train.py with three arguments: tfrecord_folder, logdir and steps. tfrecord_folder is the directory where you put your generated tfrecord, logdir is the place where you want to save the training information (models, losses, etc), and steps is the number of steps in training process. Here is an example:

python inception_2d_train.py /mnt/sdc/dataset/tf/2d /mnt/sdc/dataset/tmp/2d_inception_model 50000

Then you should see something like this:

INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Saving checkpoints for 0 into /mnt/sdc/dataset/tmp/2d_inception_model/model.ckpt.
INFO:tensorflow:loss = 3.7707667, step = 1
INFO:tensorflow:accuracy = 0.03125, global_step = 1, loss = 3.7707667, my_acc = 0.03125
...

Typically you need to wait a few hours until the training stops.

The dataset can be download from here(coming soon).

If you find this dataset is useful, please cite our paper:

3D Pyramid Pooling Network for Liver MRI Series Classification.
Zhe Zhu, Amber Mittendorf, Erin Shropshire, Brian Allen, Chad Miller, Mustafa R. Bashir, and Maciej A. Mazurowski, 
IEEE Transactions on Pattern Analysis and Machine Intelligence.

Please contact Zhe Zhu([email protected]) if you have any question about this dataset.