A generative vision transformer model for kidney tumor classification in computed tomography images
Wu Deng, Jia Hao Xu, Boyuan Ding, Yi Wei, Chao Wei, Hui Pu, Songcen Dai, Xinpeng Ren · Alexandria Engineering Journal · 2025
The accuracy of diagnosis plays a crucial role in the treatment and prognosis of kidney cancer. A high-precision kidney tumor classification model based on deep learning can help patients receive accurate diagnostic results more promptly, thereby facilitating subsequent treatment. However, traditional convolutional neural network (CNN)-based deep learning models struggle to effectively process kidney computed tomography images. This is due to individual variability in kidney morphology and the similarity in grayscale values between kidneys and neighboring organs such as the liver and spleen in omputed tomography images, which can lead to misclassification by CNN-based models. Furthermore, training traditional CNN-based kidney tumor classification models typically requires a large amount of high-quality labeled data. To address these challenges, we propose a Generative Vision Transformer (ViT) model, named GVITKT, for kidney tumor classification. This model tackles the issues of long-range dependencies and data sparsity. By leveraging the advantages of both the Vision Transformer and Generative Adversarial Networks (GAN), our approach overcomes limitations inherent in existing models. Compared to CNN, the ViT architecture is more effective at capturing global dependencies in kidney omputed tomography images, providing significant improvements in feature extraction. In addition, GAN can generate high-quality synthetic kidney tumor omputed tomography images, increasing the quantity and diversity of training data and thereby enhancing the model’s generalization to unseen data. Experimental results on a real kidney cancer omputed tomography image dataset show that the GVITKT model achieves scores of 0.999 in precision, 0.997 in accuracy, and 0.995 in F1-score, outperforming other comparison models.