RenSeg: leveraging unsupervised segmentation using localization and contour-guided quickshift for renal calculi and carcinoma segmentation and classification

Abstract

Chronic kidney diseases present a major global health challenge, often leading to renal failure due to delayed diagnosis and the shortage of nephrologists, which limits timely treatment. Therefore, the development of AI-driven systems for early detection is essential. However, most existing automated renal disease detection methods rely on supervised learning frameworks that require large amounts of clinically annotated data, which are often scarce and difficult to obtain. To address these challenges, this study proposed kidney localization and an unsupervised contour-guided quickshift-based automated segmentation method namely RenSeg on a dataset of 8,737 axial and coronal CT-scan images. RenSeg focuses on two critical diseases: Renal Calculi, which are highly prevalent and associated with recurrent pain and long-term kidney dysfunction, and Renal Carcinoma, a life-threatening tumor if not detected early. The analysis shows that unsupervised segmented images outperform manually annotated datasets. RenSeg achieved Dice scores of 0.9458 for calculi and 0.9309 for carcinoma, a precision of 0.95 and a recall of 0.94. MobileNetV2 reached the highest classification accuracy of 0.98 on RenSeg compared to 0.92 on manual annotations. The proposed unsupervised RenSeg approach outperformed the manually annotated approach across all three models in terms of accuracy, F1-score, and ROC-AUC, demonstrating superior generalization. This research demonstrates that optimizing the region of interest (ROI) enhances predictive accuracy while reducing annotation effort, making RenSeg a scalable solution for timely renal disease detection.