QMTNet

qMTNet: Accelerated Quantitative Magnetization Transfer Imaging with Artificial Neural Networks

This is the accompanying codes and data for the paper qMTNet: Accelerated Quantitative Magnetization Transfer Imaging with Artificial Neural Networks.

Overview

Figure 1 Figure 2
qMTNet is a family of neural networks that seek to accelerate qMT acquisition and fitting. qMTNet-1 (Figure 2c) is a single network that directly maps undersampled MT images to qMT parameters. qMTNet-2 composes of 2 networks, qMTNet-acq (Figure 2a) that produces 8 MT images from 4 MT images and qMTNet-fit (Figure 2b) that generate qMT parameters from 12 MT images. Both networks can produce high quality quantitative maps with a fraction of the processing time compared to conventional fitting methods.

Requirements

The codes in this repository has been developed and tested in Ubuntu 16.04.6 and Windows 10 with Anaconda. Versions of notable packages are as followed.

Python				3.5.2
Tensorflow			1.15.0
Keras				2.3.1
Numpy				1.18.1

Data

Samples of the data used in this study can be found in the data folder. This data was acquired with a Siemens Tim Trio 3T scanner.

• sample_train_data.mat:

  • Contains input, output data for pre-saturation and inter-slice MT data
  • Converted from 2D slice to pixelwise data
  • Data format: [Number of pixels] x [14 features]. The 14 features are T1, T2, and 12 MT intensities of the pixel.
  • Non-valid pixels (background) has been removed.

• sample_test_data.mat:

  • Same as sample_train_data.mat but non-valid pixels are kept so you can use reshape to get original slice.

Model

Sampled of the trained models can be found in the models folder. The model name denotes the type of model (qMTNet-fit, qMTNet-acq, or qMTNet-1) as well as the type of data that the model was trained on (conventional or inter-slice). The models included here has not been trained on the data found in the data folder.

Codes

Source codes for training and testing the models are described briefly below. You can find more details by looking through the comments inside the codes.

• model.py: define various models that was used in this study. You can define your own model here.
• utils.py: contain utility function to visualize the output of the network and the ground truth.
• train.py: train the model
• test.py : test the model

To train the network with the sampled data, you can use this command:

python train.py --data_dir ./data/sample_train_data.mat --data_mode conv --name example_exp --gpu_ids 0 --checkpoints_dir ./checkpoints --model_type qMTNet_fit

To test the included model:

python test.py --data_dir ./data/sample_test_data.mat --data_mode conv --name example_exp --gpu_ids 0 --checkpoints_dir ./checkpoints --model_type qMTNet_fit --model_dir ./models/qMTNet_fit_conv.h5

Reference

To be added