Neural networks for semantic segmentation of UK Biobank neck-to-knee body MRI

This repository contains PyTorch code for cross-validation and inference with neural networks for kidney segmentation on UK Biobank neck-to-knee body MRI, as described in:
"Kidney segmentation in neck-to-knee body MRI of 40,000 UK Biobank participants" [1]

The included inference pipeline and trained snapshot enables measurements of left and right parenchymal kidney volumes (excluding cysts and vessels) from these images.

Contents:

• 2.5D U-Net architecture with residual connections (based on TernausNet, Iglovikov et al. 2018)
• Infrastructure for training and cross-validation
• Pipeline for inference on neck-to-knee body MRI DICOMs
• Code for quality_controls based on numerical metrics
• A trained snapshot for parenchymal kidney tissue can be found here

For any questions and suggestions, feel free to reach out!

Notes

Access to the underlying image data can only be granted by the UK Biobank Study. Annotations from the quality controls used in our work [1] are available under return data ID 2345 for our application 14237. Our measurements and annotations for the kidneys will eventually be made available as well.

Citation

If you use this code for any derived work, please consider citing [1] and linking this GitHub.

References

[1] T. Langner, A. Östling, L. Maldonis, A. Karlsson, D. Olmo, D. Lindgren, A. Wallin, L. Lundin, R. Strand, H. Ahlström, J. Kullberg, “Kidney segmentation in neck-to-knee body MRI of 40,000 UK Biobank participants,” Scientific reports 10.1 (2020): 1-10\