Automatic diagnosis of ureteral stone and degree of hydronephrosis with proposed convolutional neural network, RelieF, and gradient-weighted class activation mapping based deep hybrid model

Urinary system stone disease is a common disease group all over the world. Ureteral stones constitute 20% of all urinary system stones. Ureteral stones are important because they can cause hydronephrosis and related renal parenchymal damage in the kidneys. In the study, a hybrid model was developed to detect hydronephrosis and ureteral stones from kidney images. In the developed model, heat maps of the original images were obtained by using gradient-weighted class activation mapping (Grad-CAM) technology. Then, feature maps were extracted from both the original and heatmap datasets using the Efficientnetb0 architecture. Extracted feature maps were concatenated using a multimodal fusion technique. In this way, different features of an image are obtained. This has a positive effect on the performance of the model. The Relief dimension reduction technique was used to eliminate unnecessary features in the obtained feature map so that the proposed model can work faster and more effectively. Finally, the optimized feature map is classified in the support vector machine (SVM) classifier. To compare the performance of the proposed hybrid model, results were obtained with 8 state-of-the-art models accepted in the literature. Among these models, the highest accuracy value was achieved in the Efficientnetb0 architecture with 67.98%, whereas the accuracy of the proposed hybrid model was 91.1%. This value indicates that the proposed model can be used for HUN diagnosis.

Automated summary

Urinary system stone disease is a common global disease, with ureteral stones accounting for 20% of all urinary system stones. The developed hybrid model uses Grad-CAM technology to obtain heat maps and extract feature maps from kidney images. The features are concatenated using multimodal fusion and then reduced using Relief dimension reduction technique for faster and more effective performance. The SVM classifier achieved an accuracy of 91.1% for HUN diagnosis, outperforming state-of-the-art models.