fastMRI Image Reconstruction Challenge 2019 (Single-coil track)

|

The project is structured as follows.

Challenge description

Given an undersampled knee MRI scan, the goal is to reconstruct a high resolution knee MRI scan. More details about the dataset and task can be found here.

Our method

We processed the data at the slice level. For each knee MRI low resolution, there was a corresponding high resolution knee MRI. On this processed data, we trained a U-Net architecture with a pretrained ResNet backbone on the knee MRI slices. Refer to this notebook for code implementation.

Dependencies

This work is implemented in Python 3.6 and Keras using Tensorflow as backend.

• Ubuntu 14.04
• Python 3.6

Directory strucuture and usage

• media : Contains supporting material for README.md
• dataset : training data provided by competition
• fastMRI : fastMRI github repository for helpers and utils
• *.ipynb # notebooks and python scripts
• *.py

Dataset directory strucuture:

dataset/
    singlecoil_train/
                # *.h5 files of MRI data
    singlecoil_test_v2/
                # *h5 raw test samples
    # preprocessed
    singlecoil_train_3D_images_48x/
                                low/
                                    # undersampled 3D image volumes
                                high/
                                    # ground truth 3D  image volumes
                           

Challenge Leaderboard 2019

A total of 17 teams came into the final leaderboard, among which we were the last! Some logs are shown below.

Reference to other models

Some helper scripts are based on https://github.com/facebookresearch/fastMRI.

License

Your driver's license.