Abstract & Overview
Purpose This study investigates the application of machine learning (ML) to enhance the early detection of Chronic Kidney Disease (CKD), addressing the limitations of traditional binary classification methods by implementing a multiclass classification approach using medical imaging data. Design/methodology/approach We conducted an empirical evaluation of five deep learning models – Convolutional Neural Network (CNN), ResNet50, EfficientNetB0, Vision Transformer (ViT) and a Hybrid CNN-Transformer architecture – using a publicly available dataset comprising 12,446 CT images. These images are categorised into four classes: cyst, normal, stone and tumour. Models were trained to perform a four-way classification to predict the presence and type of CKD-related conditions. Findings Among the tested models, ResNet50 achieved the highest performance, with an F1 score of 0.781 and an accuracy of 0.8304. The findings indicate that deep learning algorithms, particularly ResNet-based architectures, offer high diagnostic precision for multiclass CKD prediction using image data. Practical implications The proposed approach offers scalable solutions for healthcare systems to automate early CKD screening, potentially reducing diagnostic delays and improving patient outcomes. The implementation of such models within clinical workflows can enhance diagnostic consistency, assist overburdened practitioners and facilitate the deployment of AI-assisted decision support tools in radiology and nephrology settings. Originality/value This study presents a novel contribution by shifting the focus from binary to multiclass CKD classification using a large-scale image dataset. It provides a comparative analysis of modern deep learning architectures and highlights their applicability in clinical decision support. The work paves the way for more nuanced diagnostic tools and offers a replicable framework for future research in chronic disease detection.