Ultrasound Image Classification for Kidney Stone Detection Using CNN-SVM Hybrid Model

Abstract

Kidney stones are a common urological problem that, if not identified and treated promptly, can result in excruciating pain and complications. This paper suggests a novel method based on classification algorithms for kidney stone detection. There are several important steps in the approach. An ultrasonic image dataset must be first collected and preprocessed with adaptive thresholding, Gaussian filter, and unsharp masking before trying to enhance the picture and remove noise. The dataset is then shortly separated into train and test sets. Data augmentation techniques are used on the training set to improve its diversity. The individual CNN architecture, which is designed for the feature extraction process from utile images. It consists of multiple convolution layers, max-pooling layers, ReLU (rectified linear unit) layers etc. and the model is trained on a train set using Adam optimizer. The ultrasound dataset is processed to extract Grey-Level Cooccurrence Matrix (GLCM) features and these are concatenated with the CNN extracted features. We concatenate the extracted features and employ a Support Vector Machine (SVM) classifier to learn. Testing: In the testing phase, accuracy as well as sensitivity and specificity is computed for the learned classifier on the test dataset. This paper provides a robust method for the detection of kidney stones based on classification algorithms. The method it proposes can help in swifter detection of stone formation in kidneys than its changes and detection for enhanced patient outcomes.

Obtained results via the proposed method reached 93.22%, with a sensitivity of 92,5% and specificity of 93.59%.