Complex-valued classification of Knee pathologies

The implementation of this code uses the paper "Importance of Alternate Signal Representation in Automatic MR Image Diagnosis" [1] as baseline and creates a complex-valued deep network for the classification of knee pathologies.

Abstract

Magnetic Resonance imaging (MRI) is a commonly used medical imaging method that acquires MR signals represented by k-space which contains complex-valued frequency space measurements. The visually readable images are then obtained using the inverse Fourier Transform of the magnitude of these complex measurements, and analyzed by the physi- cians to establish a medical diagnosis. There have been numerous studies on the automation of medical diagnosis using magnitude-based image datasets as input to Deep Learning mod- els, which demonstrated very encouraging results. However, in doing so, the images chosen discard a part of the information contained in the complex-valued MR measurements no- tably in the phase of the signal. Some work has also highlighted the improvement of such results when using real-valued alternate signal representations with real-valued deep net- works. Therefore, this paper explores a novel approach that takes into account the phase information by using a raw complex-valued k-space dataset with a complex-valued net- work and focuses on comparing the classification performance of real-valued deep neural networks with real-valued alternate signal representations.

Installations

• Requirements: to install the required dependencies, run the following:

$ pip install -r requirements.txt

For data generation and processing, start by following the same steps as in paper [1] GitHub:

• Data Generation

The fastMRI can be downloaded from this link and the annotations can be found at this link. The original fastMRI data contains volume level slices so generate the Slice level processed data after updating the required paths in the file:

$ cd data_generation
$ python knee_singlecoil.py

• Generate the train, validation, and test splits

$ cd data_generation
$ python generate_knee_metadata.py

To standardize the data on the full training set, after the data generation process, run the following:

$ cd data_generation
$ python scale_complex_data.py

To train the model, run the script (after adding the appropriate argparse arguments):

$ cd src
$ python model_training.py

This model can be trained using scaled and unscaled data, 4 different activation functions, the type of output after the final complex layer.