Clinically Oriented Deep Learning Framework for Multi-Class Kidney Abnormality Instance Segmentation in CT Images

This study presents a clinically oriented deep learning framework for multi-class kidney abnormality instance segmentation in Computed Tomography (CT) images using the YOLOv12 architecture. A comprehensive and class-balanced dataset of 14,761 axial and coronal CT scans was curated, preprocessed, and augmented to support robust model training and generalization across cyst, stone, and tumor cases. The YOLOv12 models were trained and evaluated separately for axial and coronal orientations, demonstrating strong performance in both views, with mAP@0.5 scores of 0.946 and 0.885, respectively. Training analysis showed rapid convergence in classification tasks, nearly identical segmentation performance across orientations, and slightly superior bounding box localization in axial scans due to clearer spatial cues. Model testing on high-quality and simulated low-quality images confirmed consistent kidney detection (0.92–0.96 confidence) and reliable abnormality identification under challenging imaging conditions, emphasizing clinical robustness. The final system was deployed via a Streamlit-based web interface enabling real-time, privacy-preserving inference with intuitive visualization of detected organs and abnormalities. The results demonstrate the feasibility and clinical promise of YOLOv12 for supporting faster, more consistent kidney abnormality assessment, reducing radiologist workload, and enabling potential deployment in both well-resourced and resource-limited clinical environments.