Poly-CNN model recommendation and performance comparison for kidney disease detection

In recent years, rapidly increasing kidney diseases, especially in developed countries, have become one of the important problems that negatively affect human health. PACS (Picture Archiving and Communication System) imaging, as well as blood and urine tests, are also common to evaluate kidney function. Great success is achieved in medical image analysis with developing artificial intelligence technologies. Artificial intelligence-supported diagnoses allow physicians and pathologists to avoid potential errors, minimizing the risks caused by fatigue, carelessness and inexperience. These methods, which increase the success rate in disease diagnosis, make significant contributions to public health and early treatment. The classification process is of great importance in the diagnosis of kidney diseases. This study focused on the classification of kidney diseases using the data set obtained from the Kaggle platform. This dataset contains a total of 12,446 unique data, 2,283 of which are classified as tumor, 5,077 as normal, 1,377 as stone, and 3,709 as cyst. In the study, 70% of our data set is reserved for training, 15% for validation and 15% for testing. The performances of deep learning methods such as ANN (Artificial Neural Network), AlexNet, VGG16, VGG19, Classic CNN (Convolutional Neural Networks) and Poly-CNN proposed in this study were compared and analyzed. Extra pooling layer and connection layer have been added to the classical CNN structure to provide more stable learning. To prevent these added layers from causing excessive learning, random neurons were disabled during training. The complexity matrices and accuracy and loss graphs obtained from the parameter values, layer structures and validation data used in deep learning models were analyzed in detail. When the statistical analysis results of the architectures are evaluated, the best accuracy rate of the test data is as follows, in order of success: Poly-CNN (99.94%), VGG19 (99.78%), ANN (99.78%), Classic CNN (99%). 57), VGG16 (96.57%) and AlexNet (89.44%) results were obtained. It is thought that the study results can provide valuable suggestions to field experts and physicians in their decision-making processes. It offers advantages such as shortening the diagnosis time of the disease and minimizing human errors.