A Deep Learning Approach for Automated Kidney Tumor Classification

Abstract

Correct classification of kidney tumors in computed tomography (CT) images is crucial for early diagnosis and treatment planning. In this study, we propose a hybrid architecture in which ResNet50 is employed for localized spatial feature extraction, while Vision Transformer (ViT) enables global contextual learning to automatically classify kidney tumors into multiple classes. A single-stage paradigm was then performed to classify CT images into one of the three clinically relevant categories: normal tissue, benign tumor and malignant tumor. All classifications and evaluations were performed at the slice level, where each axial CT slice was treated as an independent input sample. The model was independently trained and tested using two publicly available datasets of CT images, KAUH-Kidney, and CT-Kidney. It so adapts advanced preprocessing methods like class-aware data augmentation, normalization, and focal loss to deal with class imbalance. To thoroughly evaluate performance, we used a 5-fold cross-validation. On the KAUH-Kidney dataset, the hybrid model reached an accuracy of 99.53% and on the CT-Kidney dataset it detected 99.73% of the samples with the perfect macro and micro AUC scores on both datasets. Where the combined approach outperformed both standalone CNN or transformer-based architectures in accuracy, F1-score, and generalization. This study demonstrates the potential of hybrid deep learning frameworks to assist in the more accurate, efficient and automated classification of kidney tumors in clinical practice settings