Renal-Net: Renal mass segmentation

---

About

Summary

This algorithm runs Renal-Net, a segmentation model that segments the kidney region and renal masses (including tumours and cysts) in contrast-enhanced CT scans. It was developed at the Diagnostic Image Analysis Group, Radboudumc, Nijmegen, as part of the OncoAI research group.

Important: Please note that the algorithm that can be tried here is a 'fast mode' version, meaning it only uses one fold and less test time augmentation. With a small test set, we only noticed small (not significant) performance differences.

Method. The pipeline consists of three stages. First, an optional preprocessing step uses TotalSegmentator to localise the lung lobes and urinary bladder and crop the scan to the abdominal region of interest, reducing memory usage for large volumes. Second, a residual-encoder nnU-Net (nnUNetResEncUNetLPlans, 3D full-resolution, 5-fold ensemble) segments the cropped volume into background, kidney (label 1), and renal mass (label 2). Third, postprocessing removes abnormality components smaller than 3 mm in diameter and retains only components that are spatially connected to a kidney region; large components exceeding 100 mm³ are kept regardless of connectivity.

Training. The model was trained on a combination of KiTS23 training data and a publicly available Radboudumc kidney dataset of contrast-enhanced CT scans with annotations of kidneys and kidney abnormalities.

A validation study about this model is published in MELBA and can be found here: paper.

Mechanism

Target population. Adult oncology patients undergoing contrast-enhanced CT for renal mass evaluation, staging, or follow-up. The algorithm is designed to handle the full range of kidney morphologies, including patients with a single kidney, bilateral tumours, and large or complex masses.

Algorithm description. The algorithm follows a three-stage pipeline:

1. ROI extraction (optional). When enabled via the crop-roi flag, TotalSegmentator (fast mode) is run on a downsampled version of the CT (3 mm isotropic) to localise the lung lobes and urinary bladder. The resulting bounding box is mapped back to the original resolution and used to crop the scan to the abdomen, reducing peak memory consumption.

2. Segmentation. The (cropped) CT is resampled to 0.75 mm isotropic spacing and passed to a residual-encoder nnU-Net (nnUNetTrainer, nnUNetResEncUNetLPlans, 3D full-resolution) trained on public data. Predictions from all five cross-validation folds are ensembled using Gaussian-weighted patch aggregation.

3. Postprocessing. Kidney abnormality components smaller than 3 mm in maximum axial diameter are discarded. Remaining components are retained only if they overlap with a 4 mm dilation of the predicted kidney mask. Components exceeding 100 mm³ are always retained regardless of kidney proximity.

Inputs. The algorithm accepts a single CT scan in MHA format and a boolean crop-roi flag. When crop-roi is false, the full scan is passed directly to the segmentation network.

Outputs. A single-channel segmentation mask in MHA format, resampled to the original CT spacing and field of view. Label convention: 0 = background, 1 = kidney, 2 = renal mass.

---

Interfaces

This algorithm implements all of the following input-output combinations:

Validation and Performance

All results report mean ± standard deviation. Results are for Renal-Net without ROI cropping, and with post-processing enabled.

TCGA-KIRC

Region	Dice ↑	HD95 (mm) ↓
Kidney	0.95 ± 0.03	3.79 ± 5.89
Kidney + mass	0.97 ± 0.02	3.64 ± 7.85
Renal mass	0.86 ± 0.25	14.94 ± 25.81

Charité Universitätsmedizin Berlin

Region	Dice ↑	HD95 (mm) ↓
Kidney	0.85 ± 0.10	9.84 ± 16.11
Kidney + mass	0.90 ± 0.08	9.21 ± 11.92
Renal mass	0.83 ± 0.17	10.46 ± 15.83

Radboudumc — B20: 20 patients with healthy kidneys; B30: 30 patients with renal lesions.

Region	Dice B20 ↑	HD95 B20 (mm) ↓	Dice B30 ↑	HD95 B30 (mm) ↓
Kidney	0.93 ± 0.06	55.80 ± 101.21	0.94 ± 0.05	44.32 ± 99.98
Kidney + mass	—	—	0.95 ± 0.02	43.66 ± 100.00
Renal mass	—	—	0.66 ± 0.29	31.84 ± 39.10

Uses and Directions

Intended use. This algorithm is intended for research purposes only. It is not a certified medical device and must not be used for clinical decision-making, diagnosis, or treatment planning.

The algorithm is designed to support research workflows involving automated segmentation of kidneys and renal masses in contrast-enhanced CT scans. Potential research applications include cohort-level analysis, lesion quantification, and the development or evaluation of downstream AI models.

Limitations. - Performance may degrade on CT scans acquired with non-standard protocols (e.g. non-contrast, MRI, or low-dose scans). - The algorithm was validated on adult patients; performance on paediatric populations is unknown. - Very small lesions (< 3 mm) are removed by postprocessing and will not appear in the output. - Predictions should not be used as a substitute for expert radiological review.

Not intended for. - Clinical diagnosis or treatment decisions. - Real-time or intraoperative use. - Any regulatory submission without independent clinical validation.

Warnings

• Not for clinical use. This algorithm is a research tool and has not been validated for clinical decision-making. It must not be used to guide diagnosis, staging, treatment planning, or any other clinical workflow.

• No regulatory approval. This algorithm has not been approved or cleared by any regulatory body (e.g. CE, FDA).

• Imperfect predictions. All outputs are probabilistic and may contain false positives, false negatives, or inaccurate boundaries. Renal mass segmentation in particular showed higher variability (Dice 0.66–0.86 across test sets).

• Out-of-distribution scans. The algorithm was trained and validated on contrast-enhanced CT. Performance on non-contrast CT, MRI, PET-CT, or low-quality or artefact-heavy scans is unknown and may be substantially degraded.

• Small lesion detection. Components smaller than 3 mm in diameter are removed during postprocessing and will be absent from the output mask.

• Patient population. Validation was performed on adult oncology patients. Applicability to paediatric patients has not been established.

Common Error Messages

Error message	Likely cause	Solution
No CT files found under /input/images/ct	No .mha file placed in the expected input directory, or file has a different extension.	Ensure the input CT is in .mha format and placed at /input/images/ct/<filename>.mha.
Missing crop-roi.json at /input/crop-roi.json	The crop-roi boolean interface value was not provided.	Ensure the crop-roi input is connected and set to true or false in the algorithm interface.
[run] ROI extraction failed (...), falling back to segmenting the full CT.	TotalSegmentator could not locate the lung lobes or urinary bladder, typically because the scan does not cover the full abdomen.	This is handled automatically — the algorithm will continue segmenting the full scan. No action required.
Failed during segmentation and postprocessing	Insufficient GPU memory for the input volume, or a corrupt input file.	If the scan is very large, enable crop-roi to reduce memory usage. Check that the input file is a valid CT scan and not corrupted.
Algorithm times out without producing output	Input volume is very large and inference exceeds the time limit.	Enable crop-roi to crop the scan to the abdominal region before segmentation, significantly reducing runtime.