yanielc / textural-bias-medical-imaging

Working with textural filters and DCNNs.

Textural bias medical imaging

MRI artifacts

During data acquistion and image reconstruction different types of artifacts can make their way into the final MR images. In this repository I have implemented transforms that model three such artifacts: Gibbs (or truncation) artifact, spikes (herringbone) artifact, and wraparound (alasing) artifact. Each transform is implement in the k-space version of the data.

More details on the physics of the artifacts can be found in the following publications:

AAPM/RSNA physics tutorial for residents: fundamental physics of MR imaging.

Body MRI artifacts in clinical practice: A physicist's and radiologist's perspective.

An Image-based Approach to Understanding the Physics of MR Artifacts.

Examples and visualizations

For an overview of the three transforms and visualizations of their applications on the data refer to the Jupyter notebook: artifacts_transforms_visualizations.ipynb.

MONAI contributions

I have contributed the Gibbs filter to the codebase of MONAI as various transfroms:

1. Related to Gibbs artifacts:

   • GibbsNoise: applies artifact directly on the image.
   • RandGibbsNoise: to apply randomly on images with uniform sampling of filter's intensity.
   • GibbsNoised: to apply on group data; dictionary-style version of GibbsNoise.
   • RandGibbsNoised: dictionary-style of RandGibbsNoise.

2. Related to Spikes (herringbone) artifacts:

   • KSpaceSpikeNoise: applies spike artifacts directly on the image.
   • RandKSpaceSpikeNoise: applies the artifacts randomly on images while sampling in the intensity of the artifacts.
   • KSpaceSpikeNoised: allows to apply transform both on the image and/or label; dictionary-style version of KSpaceSpikeNoise.
   • RandKSpaceSpikeNoised: dictionary-version of RandKSpaceSpikeNoise.

Textural filters and DCNN

• Segmentation experiments are performed using a residual Unet.

Robustness to the presence of artifacts across a range of intensities.

• Training the models on stylized data results in improved performance when it comes to unseen data with different strengths of the artifacts.
• Examples of this results are shown in the notebook stylized_models_inference.ipynb.

Robustness to unseen sourcing distribution.

• Early experiments are showing that stylized data are more robust when tested on images coming from hospitals unseen at training time.

Still in the works: compressed MRI sensing.

• Reconstruction problem in the presence of undersampled data.
• Network: ReconGan.