Automated Renal Analysis

The problem

Renal measurement is a routine part of abdominal imaging, but it is difficult to automate reliably from thick-slice non-contrast CT. With slices spaced 8 mm apart, the voxels are highly anisotropic. The kidneys are also angled rather than aligned with the scanner axes, so measuring across axial slices captures a projection of the kidney rather than its true longitudinal extent.

Approach

The kidneys are first segmented using a pretrained model. Renal length is then calculated from the principal-axis extent of the kidney mask in world-millimetre coordinates. This allows the measurement to follow the kidney’s own orientation rather than the scanner’s axes. The final measurement is reported in millimetres using the image affine, rather than as a number of slices.

The system is designed as a broader renal analysis pipeline, including renal measurement, stone and cyst segmentation, and automated report generation from thick-slice CT.

The dataset is part of the contribution

The dataset is de-identified, with each case linking a volumetric NIfTI scan to its DICOM-derived metadata and radiologist report. It includes multi-label annotations for urolithiasis, hydronephrosis, renal cysts, pyelonephritis, cystitis, nephrocalcinosis, renal masses, and normal studies.

Radiologist reports are also used as the reference source for renal measurements and automated report generation.

Validation

The validation cohort was deliberately stratified to include more cases where renal measurement is particularly challenging and clinically relevant, including hydronephrosis, staghorn calculi, and atrophic kidneys.

Status

In preparation. Results are held until publication.