Optimized TSA ResNet Architecture with TSH—Discriminatory Features for Kidney Stone Classification from QUS Images

Kidney diseases are the major reason for renal failure. Ranging from calcium deposits, stones, and to the maximum extent of chronic kidney disease, there are multiple classifications of that which may cause renal failure and lead to a large proportion of mortality. Qualitative Ultrasound images are usually preferred as the ground for examining the kidney in medical contexts. In recent times Computer-Aided Diagnosis of kidney health analysis has paved the way for the effective detection of diseases at early stages by employing convolutional Neural Networks and their allied versions of deep learning technologies. The availability of these algorithms in a simulated environment yields better results when compared to images taken in real-time cases. The performance of these algorithms is confined within a limited level of performance metrics such as accuracy and sensitivity. To address these issues, we have focussed on building an automated diagnosis of kidney diseases and classifying it according to their features illustrated in the QUS images. The anticipated methodology in this work merges the texture, statistical and histogram-based features (TSH) which are discriminative when compared with other features exhibited by the QUS, then these TSH features are employed in ResNet architecture for successful recognition of kidney diseases. The observance in the reduction of accuracy due to the improper training of the hyperparameters such as momentum and learning rate of CNN is obliterated with the usage of the position-based optimization algorithm, namely the Tree Seed Algorithm. The output of the classification was analysed through the performance analysis for the optimization-tuned kidney image standard dataset. The results from the ResNet model with TSA optimization show quite good efficiency of using an algorithmic approach in tuning deep learning architectures. Further exploration of the momentum and learning rate of the Resnet architecture makes the proposed TSH-TSA-Resnet architecture outperform the existing method and provide a classification accuracy of 98.9%.