SRENet: a spatiotemporal relationship-enhanced 2D-CNN-based framework for staging and segmentation of kidney cancer using CT images

Kidney cancer (KC) is among the 10 most common cancers posing health threats to humans, with an average lifetime risk of 1.53%. Computed tomography (CT) is regarded as the golden standard for the characterization of KC and is widely used for KC prognosis. However, it is challenging to segment KC in CT images and perform cancer staging simultaneously due to the variable positions and shapes of kidney tumors and similar textural features in the background and target areas. We propose a novel spatiotemporal relationship-enhanced convolutional neural network (CNN)-based framework called SRENet. It consists of a spatial transformer framework and a residual U-Net with a temporal relationship extraction module for the staging and segmentation of KC. The SRENet achieves excellent performance on six evaluation metrics (Kappa, Sensitivity, Specificity, Precision, Accuracy, F1-score) for KC staging with an F1-score of 98.46%. The framework demonstrates a strong and reliable capacity for kidney and KC segmentation with Dice coefficients (DCs) of 97.89% and 92.54%, respectively, outperforming the state-of-the-art models (ResNeXt-101, ViT and Swin-Transformer for staging, and VNet, UNet 3+ and nnUNet for segmentation). The proposed SRENet helps accelerate the development of reliable kidney and KC segmentation methodologies and shows significant potential for KC diagnosis in clinical practice.