SahiKid: An Efficient SAHI-based Framework for Small Kidney Stone Detection in CT Scans

Kidney stones are a prevalent urological issue, and early detection—particularly of small stones on CT scans—is critical for effective treatment and the prevention of complications. However, accurately identifying these tiny structures remains challenging due to their size and the complexity of surrounding anatomy. This study proposes an efficient detection framework that integrates a lightweight object detection (OD) model (i.e., Yolov11n) with Slicing Aided Hyper Inference (SAHI) techniques. In addition, we introduce a novel algorithm, termed B-SAHI, to improve computation time. Our proposals are evaluated on a public CT dataset under various settings, including full-image inference (FI), Slicing Aided Fine-tuning (SF), patch overlapping (PO), and kidney region localization (KD). The best-performing implementation—OD with SF, SAHI, PO, and KD—achieved a mean Average Precision (mAP50) of 78.66%, vastly outperforming the 0.05% baseline. In terms of efficiency, B-SAHI reduced the computation time from 0.1553 to 0.0329 seconds, enabling real-time performance. These results underscore the effectiveness of slice-based inference and anatomical focus in enhancing small object detection in medical imaging.