Skip to content
KidneyScanAI

About KidneyScan AI

KidneyScan AI is a research prototype from Stochastic Research Labs for exploring automated kidney tumor and cyst segmentation on CT scans. It is meant for students learning CT segmentation and for small teams without a kidney specialist.

The data

KiTS23 provides 599 cases: 489 public training cases and 110 test cases with private masks. Every voxel is labelled background, kidney, tumor or cyst. The data is licensed CC BY-NC-SA 4.0: free for research, not commercial. It is not stored in this project.

The model

A light 3D U-Net with 1.68M parameters reads a whole CT volume and labels each voxel. It is far smaller than the ensembles used in challenge entries, which is the point: it is meant to be easy to run and easy to study. The results are modest, and the model insights page shows them as they are.

The viewer

Your scan is read in the browser and drawn with Three.js by ray-marching through the volume. The original CT and the segmentation overlay share one camera, so you can rotate a full 360 degrees and compare the same angle. Volumes in cm³ are counted from the predicted mask using the voxel size stored in the scan.

Limits

This is not a medical device. Predictions can be wrong, especially for tumors and cysts, and the model gives no per-case accuracy. A doctor must confirm any finding. The home page kidney is an illustration generated in code, not a scan or a result.