Enhancing kidney stone segmentation in computed tomography images through successive scope reduction and intra-framework feature transfer

Kidney stone diseases (KSDs) are one of the oldest documented renal conditions, with cases tracing back to ancient history. Due to its recurrent nature, KSDs often require patients to undergo lifelong periodic monitoring. Artificial intelligence can automate and speed up the tracking of kidney stone presence, growth, and location. Computed tomography (CT) remains one of the most widely used methods for KSD diagnosis. However, accurately segmenting kidney stones within a full abdominal CT scan is challenging because of their small size, necessitating extensive scope reduction to enable the deep learning (DL) model to focus on the relevant regions. While traditional saliency-map-based localization provides a general estimate of stone location, it lacks precise details on the stone’s size, shape, and specific position within the renal cortex. Addressing these limitations, this study introduces a dual-stage, DL-based framework for efficient kidney stone segmentation from CT slices. First, the dataset scope is reduced using the proposed successive scope reduction (SSR) method, sequentially narrowing the kidney and stone focus for enhanced accuracy. With the successive scope reduction network (SSRNet) classification model, an overall F1-score of 94.45% is achieved for scope reduction. On the scope-reduced dataset, the kidney stone segmentation network (KSSNet) model reaches a dice score of 97.44% and a Jaccard index of 95.01% for stone segmentation, significantly surpassing existing benchmarks. The proposed framework is robust, with subject-independent evaluations and ablation studies confirming its reliability and computational efficiency, and the proposed models surpassing the state-of-the-art in respective domains, supporting its application in a real-world clinical setup.