Automated Kidney Stone Detection using Deep Learning Technique

Convolutional Neural Networks(CNN) are created to work mostly on the image datasets and have revolutionized image classification and object detection by introducing versatile architectures which can be modified according to the requirements needed with help of modifiable hyperparameters like architecture specifications, batch size, kernel, stride size, loss function, learning rate etc.,. The use of ResNet architecture introduces residual pathways which accelerates weight convergence compared to the traditional neural networks and other CNN architectures like AlexNet, LeNet, GoogLeNet by effectively preserves much patterns and informations contained in the images and giving a good accuracy in almost all cases considering the dataset size, quality and task complexity. In this proposed work, the CT images of Kidneys which are divided into train (1453 images) and test (346 images), encompassing both stones and non-stone cases. Employing ResNet50 architecture with meticulously configured hyperparameters and tailored preprocessing methods are made to learn the train data with a specific number of epochs and suggested learning rate. After training the model for 50 epochs, applied to detect stone's presence in the test set and achieved an accuracy of 93%. The limitations of conventional machine learning models in image classification tasks by considering Support Vector Machine, Logistic Regression and RandomForest demonstrates their challenges in capturing complex image patterns and features, often results in lower accuracy. And this deep learning model must need Graphics Processing Unit (GPU) for training to reduce the computation time and memory management.