Multi-Phase CNN-ViT-Wavelet Fusion with Attention for Robust Kidney Stone Detection from Ultrasound Images

Abstract

Accurate detection of kidney stones in ultrasound images is hindered by low contrast, speckle noise, and operator-dependent variability. The objective of this study is to develop a robust multi-phase model that integrates the local, global, and frequency-domain features to enable a reliable and generalizable kidney stone classification. Three convolutional neural networks (ResNet50, DenseNet121, and EfficientNet-B0) and vision transformer variants (ViT-Base, Swin-Transformer, and DeiT-Small) were independently trained on a primary renal ultrasound dataset. The dataset was split patient-wise in a 70-15-15 manner; extensive data augmentation, normalization, and five-fold cross-validation were implemented to avoid data leakage and ensure the model's robustness. The best-performing CNN and ViT features were fused at the feature level with an attention mechanism and classified using ML classifiers, among which XGBoost demonstrated the optimal performance. Next, a discrete wavelet transform (DWT) branch was incorporated to acquire complementary frequency information for further enhancing the discriminative capability. The multi-phase framework achieved 97.9%, 97.8%, and 0.997 for accuracy, F1-score, and AUC, respectively, on the internal dataset. Similarly, it obtained 94.3%, 94.0%, and 0.970 for accuracy, F1-score, and AUC on the external renal ultrasound dataset. These results demonstrate a robust generalization backed up by Grad-CAM and attention maps. The Multi-Phase Framework (MPF) offers a consistent, generalizable, and fully automated method for kidney stone detection in ultrasound images, supporting improved diagnostic performance.