ColorNephroNet: Kidney Tumour Malignancy Prediction Using Medical Image Colourisation

Motivation

Many nephrectomies (kidney removals) are performed for tumours later found to be benign. We set out to build a system that could support radiological diagnosis by predicting the malignancy of a renal tumour directly from CT images. Conventional CNN models take grayscale CT images as input; we asked whether colourising the medical images into pseudo-RGB could improve classifier performance by allowing the model 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. The kidney and tumour region was cropped from each image, intensity values were standardised from the Hounsfield scale to [0,1], images were resized to a fixed resolution, and random flips, rotations, and scaling were applied for augmentation — all with no patient overlap between train, validation, and test splits.

Method

The pipeline runs in two stages. First, an encoder–decoder CNN converts a single-channel CT crop into a three-channel colourised image using an L2 reconstruction loss, introducing richer texture and contrast cues that enable transfer from pre-trained RGB CNNs. Second, a CNN classifier (initialised from ImageNet) predicts a binary benign/malignant label from the colourised image using binary cross-entropy and early stopping.

Results

ColorNephroNet achieved an F1-score of 0.849 versus 0.831 for the grayscale baseline — an improvement of roughly 1.8 percentage points, alongside gains in accuracy (0.840 vs 0.823), precision, and recall. Models were implemented in PyTorch with the Adam optimizer, a 1e-4 learning rate, batch size 32, and weight decay 1e-5.