Precise Kidney Stone Localization in Medical Imaging via a Capsule Network

Abstract

Traditional convolutional neural networks (CNNs) face significant limitations in medical imaging when detecting small, spatially variable objects such as kidney stones, primarily due to their inability to preserve pose information and spatial correlations through max pooling operations. While previous CNN-based studies achieved approximately 93% accuracy in kidney stone detection, they struggled with the precise localization of small or partially obscured stones, creating a critical research gap in automated urological diagnostics. This study develops and evaluates a capsule network (CapsNet) framework that leverages dynamic routing and vector-based capsules to increase kidney stone localization accuracy in computed tomography (CT) images while maintaining spatial coherence and reducing false positives. The CapsNet model incorporates convolutional layers, primary capsules, and stone capsules via dynamic routing algorithms. The approach was systematically evaluated via a publicly accessible kidney stone CT dataset from the Mendeley repository, comprising 512 anonymized abdominal CT slices preprocessed to 256×256 pixels. The dataset was partitioned into training (70%), validation (15%), and test (15%) sets. The performance was compared against that of a baseline CNN under identical conditions using 50 epochs and the Adam optimizer. The results demonstrate CapsNet's superior performance across all the metrics: 96.5% accuracy, 96% precision, 97% recall, 96% F1 score, 0.93 Dice coefficient, and 0.89 IoU, significantly outperforming the CNN baseline (92% accuracy, 0.84 Dice coefficient, 0.78 IoU). CapsNets enhance kidney stone localization and generalization by preserving spatial and pose information, improving diagnostic accuracy in medical imaging.