ColorNephroNet: Kidney Tumour Malignancy Prediction Using Medical Image Colourisation
Renal kidney tumours classification
Many nephrectomies (kidney removals) are performed for tumours later found to be benign.
We wanted to build a system that could support radiological diagnosis by predicting the malignancy of a renal tumour directly from CT images.
Conventional CNN models for this task take grayscale CT images as input.
We asked whether colourising the medical images — converting them into pseudo-RGB — could improve the classifier’s performance by allowing it to leverage features learned from colour image datasets.
Data and Preprocessing #
We used CT scans of kidneys with tumours, each with a known benign or malignant outcome confirmed postoperatively.
- Region extraction – Cropped the kidney and tumour region from each CT image.
- Normalization – Standardised intensity values (Hounsfield scale → [0,1]).
- Resizing – All images resized to a fixed input resolution.
- Data augmentation – Random flips, rotations, and scaling to improve generalisation.
- Dataset split – Divided into train, validation, and test sets, ensuring no patient overlap.
Colourisation Network #
The first part of the pipeline is an encoder–decoder CNN that converts a grayscale tumour region into a pseudo-RGB image.
- Input: Single-channel CT crop
- Output: Three-channel colourised image
- Loss: L2 reconstruction loss
- Purpose: Introduce richer texture and contrast cues; enable transfer from pre-trained RGB CNNs
After training, this network is used to generate colourised tumour images for the classification step.
Classification Network #
We used a CNN classifier to predict tumour malignancy.
- Input: Either grayscale (baseline) or colourised image (proposed)
- Output: Binary label (benign/malignant)
- Loss: Binary cross-entropy
- Training: Standard supervised learning with early stopping
The ColorNephroNet pipeline combines both steps: CT image -> Crop tumour ROI -> Colourisation network -> CNN classifier -> Malignancy prediction
Experiments #
We compared two setups:
| Model | Input | Pretraining | Description |
|---|---|---|---|
| Baseline | Grayscale | Random init | CNN trained directly on CT crops |
| ColorNephroNet | Colourised | ImageNet | CNN trained on colourised CT crops |
Metrics #
We used Accuracy, Precision, Recall, and F1-score as evaluation metrics.
Results #
| Model | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|
| Grayscale baseline | 0.823 | 0.821 | 0.830 | 0.831 |
| ColorNephroNet | 0.840 | 0.838 | 0.841 | 0.849 |
ColorNephroNet achieved an F1-score improvement of ~1.8 percentage points over the grayscale baseline.
Confusion Matrix (Test Set) #
| True Label | Pred. Benign | Pred. Malignant |
|---|---|---|
| Benign | 43 | 6 |
| Malignant | 7 | 52 |
Implementation Details #
| Parameter | Value |
|---|---|
| Framework | PyTorch |
| Optimizer | Adam ofc |
| Learning rate | 1e-4 |
| Batch size | 32 |
| Weight decay | 1e-5 |
| GPU | NVIDIA RTX 2080 |
Summary #
- Extract tumour region from CT scan.
- Train a colourisation CNN to convert grayscale to pseudo-RGB.
- Feed colourised images to a CNN classifier for malignancy prediction.
- Compare to grayscale baseline on held-out test data.
- Report consistent performance improvement (~+1.8 pp F1).
Conclusion #
ColorNephroNet shows that medical image colourisation can act as an effective preprocessing step for CNN-based classification tasks.
Even with limited CT data, the colourisation step improves feature extraction and overall predictive performance for kidney tumour malignancy prediction.
This post summarises our paper presented at the 35th International Florida Artificial Intelligence Research Society Conference (FLAIRS-35):
A. Obuchowski, B. Klaudel, R. Karski, B. Rydziński, M. Glembin, P. Syty, P. Jasik — “ColorNephroNet: Kidney Tumor Malignancy Prediction Using Medical Image Colorization”.
Read the paper