Multi-scale feature fusion for image-level kidney stone classification on NCCT images: a cross-center generalization evaluation

Kidney stones are a common and serious urinary tract disease. Due to their small size, indistinct boundaries, and differences in imaging protocols and scanners across centers, deep learning-based image-level classification on non-contrast computed tomography (NCCT) images still faces challenges in reliably distinguishing stone-positive images from normal images and achieving robust cross-center performance. To this end, we developed a deep learning-based image-level NCCT kidney stone classification framework that encompasses dataset construction, unified preprocessing, systematic comparison of image enhancement strategies, modeling of a multi-scale feature-fusion convolutional neural network (MSF-CNN), and cross-center generalization evaluation. Within this framework, MSF-CNN enhances the representation of stone-related features by fusing multi-scale and global contextual information and introducing a channel attention mechanism. We evaluated 1,496 coronal NCCT images from 497 subjects at a single center with institutional ethics approval, and additionally used 369 images from a publicly available external dataset as an independent external test set. Seven image enhancement methods were systematically compared, and ablation experiments were conducted under a unified preprocessing pipeline. The results showed that Laplacian sharpening achieved the best performance among the evaluated enhancement methods, yielding 99.16% accuracy and 99.97% AUC on the internal test set and 97.02% accuracy and 98.50% AUC on the external test set when combined with MSF-CNN. Overall, the proposed framework achieved highly accurate and cross-center-robust image-level classification of stone-positive and normal NCCT images, providing a potential auxiliary screening tool for clinical decision support.