Model performance
Best results so far for the light 3D U-Net. These are honest numbers from a small model, and they leave plenty of room to improve.
Dice score by structure
Foreground (average of the three)0.4043
Kidney0.787
Tumor0.378
Cyst0.048
Dice measures overlap between the prediction and the expert outline. 1 is a perfect match and 0 is no overlap. The foreground score is the average of the three structure scores.
Reading the numbers
- Kidney (0.787). The clearest result. Still expect boundary errors.
- Tumor (0.378). The model finds part of the tumor. Do not rely on its outline or its volume.
- Cyst (0.048). Close to failing. Cysts are small and uncommon, and they are the hardest class. Treat any cyst result as unreliable.
Volumes shown in the workspace come from the predicted mask, so they inherit these errors.
Model and data
- Architecture
- Light 3D U-Net
- Parameters
- 1.68M
- Dataset
- KiTS23, 599 cases (489 public training, 110 test with private masks)
- Labels
- 0 background, 1 kidney, 2 tumor, 3 cyst
- Input
- One 3D CT volume (.nii.gz)