DCE-NET

Framework for estimating DCE-MRI physiological parameters.

Based on the master thesis: DCE-MRI parameter estimation using a physics-based deep learning approach

Getting Started:

To create the environment dce in anaconda, the following command can be used:

conda env create -f environment.yml

Simulations:

Executing the framework on simulations can be done using main.py.

usage: main.py [--nn] [--layers] [--lr] [--batch_size] [--attention]
               [--bidirectional] [--supervised] [--results]

optional arguments:
  --nn            neural network to use - linear / lstm / gru
  --layers        number and size of layers - linear:   neurons_layer_1 neurons_layer_2 ...
                                            - lstm/gru: hidden_dimension stacked_layers
  --lr            learning rate - float
  --batch_size    batch size - int
  --attention     option to include attention layers for lstm/gru
  --bidirectional option to include bidirectionality for lstm/gru
  --supervised    option to train on ground truth parameters
  --results       option to perform evaluation on network using different SNR (must be trained first)

More hyperparameters are stored in hyperparameters.py.

Showing Results

Results of the trained frameworks can be obtained by executing results.py.

Patient Data

To construct a dataset, create_dataset.py is used that stores DCE-MRI signals with a dimension of (Slices, Width, Height, Length).

Training on the dataset is done with sim_patients.py.

usage: sim_patients.py [--nn] [--layers] [--lr] [--batch_size] 
                       [--optim] [--dual_path] [--bidirectional] 
                       [--pretrained] [--cpu]

optional arguments:
  --nn            neural network to use - linear / lstm / gru / convlin / unet / convgru
  --layers        number and size of layers - linear:   neuron_layer_1 neuron_layer_2 ...
                                            - lstm/gru: hidden_dimension stacked_layers
                                            - convlin:  input_channels_1 input_channels_2 ...
                                            - unet:     first_input_channel depth
                                            - convgru:  hidden_dimension_1 hidden_dimension_2 ...
  --lr            learning rate - float
  --batch_size    batch size - int
  --optim         optimizer - adam / sgd / adagrad
  --dual_path     option to create a convolution linear architecture with a dual pathway
  --bidirectional option to include bidirectionality for lstm/gru/ConvGRU
  --pretrained    option to train further on a pretrained model
  --cpu           option to switch to cpu

Showing Results

Similarity

Executing similarity.py calculates the RMSE, nRMSE and SSIM compared to the input signals. The results are stored in patient_results.csv

The same arguments are used as in sim_patients.py to specify on which network the similarity is calculated.

Parameter maps

Executing visuals.py shows the parameter maps created from different specified frameworks.

Bland-Altman Plots

Executing bland_altman.py creates bland-altman plots from the specified frameworks and the non-linear least squares method.

The same arguments are used as in sim_patients.py to specify from which network the bland_altman plots are created.