Kidney Week

Abstract: PO0304

Kidney Segmentation with Deep Learning in MRI of 40,000 UK Biobank Subjects

Session Information

• Bioengineering
  October 22, 2020 | Location: On-Demand
  Abstract Time: 10:00 AM - 12:00 PM

Category: Bioengineering

• 300 Bioengineering

Authors

• Langner, Taro, Uppsala Universitet, Uppsala, Sweden

Taro Langner,

• Östling, Andreas, Uppsala Universitet, Uppsala, Sweden

Andreas Östling,

• Maldonis, Lukas, Antaros Medical AB, Mölndal, Västra Götalands län, Sweden

Lukas Maldonis,

• Karlsson, Albin Karl-Gustav, Uppsala Universitet, Uppsala, Sweden

Albin Karl-Gustav Karlsson,

• Olmo, Daniel, Uppsala Universitet, Uppsala, Sweden

Daniel Olmo,

• Lindgren, Dag, Antaros Medical AB, Mölndal, Västra Götalands län, Sweden

Dag Lindgren,

• Wallin, Andreas Hans, Antaros Medical AB, Mölndal, Västra Götalands län, Sweden

Andreas Hans Wallin,

• Lundin, Lowe, Antaros Medical AB, Mölndal, Västra Götalands län, Sweden

Lowe Lundin,

• Strand, Robin, Uppsala Universitet, Uppsala, Sweden

Robin Strand,

• Ahlström, Håkan, Uppsala Universitet, Uppsala, Sweden

Håkan Ahlström,

• Kullberg, Joel, Uppsala Universitet, Uppsala, Sweden

Joel Kullberg,

Background

Kidney volume and its association to several demographic and physiological parameters are subject of ongoing research. The UK Biobank (UKB) studies over half a million volunteers, examining blood samples, lifestyle, genetics, and body composition, including medical imaging for 100,000 participants, and 10,000 repeat scans. We have developed a system for automated kidney segmentation in 40,000 currently available MRI scans for image-based measurements of parenchymal kidney volume.

Methods

UKB neck-to-knee body MRI has been released for 40,264 participants (52% women), aged 44-82 (mean 64) years, with BMI 14-62 (mean 27). The kidneys are imaged in two 17s breath-hold stations with a Siemens 1.5T Aera device at (224 x 174 x 44) voxels of (2.23 x 2.23 x 4.5) mm. In this work, three operators marked cortex and medulla, excluding cysts, in 122 subjects (Fig a, b) for training and validation of a 2.5D U-Net with short skip connections. This neural network learned to segment axial slices.

Results

The network predictions matched the references in total kidney volume for a mean error below 4% (or 10 cm³, Dice score 0.956), whereas human repeat segmentation yielded 3% (or 6 cm³, Dice score 0.962). While imaging limitations such as motion may compound this error, similar performance is expected for future UKB releases. After 30 minutes of training, the network can process all scans within two days. Exclusion of anomalies, such as 40 cases of renal fusion, left 37,468 subjects with median voxel count volumes of 277 cm³ for men and 220 cm³ for women (Fig c).

Conclusion

The proposed system may ultimately provide measurements of left and right kidney volume for all imaged UKB subjects which can be analyzed and shared for further large-scale investigation of associations and longitudinal changes in kidney volume.

Segmented kidneys in MRI (a, b) and measured volumes in the entire cohort (c).

Funding

• Other NIH Support