Automated Kidney Segmentation and Disease Classification Using CNN-Based Models

Abstract

According to studies, kidney disease (KD) is predicted to be the 5^th leading cause of death by 2040. Some works have been done on KD segmentation and classification using CNN-based models. However, works that consider automated KD segmentation and classification using CNN-based models for more than one KD type, together with model optimization are rarely investigated. In this study, four KD conditions (tumor, cyst, stone, and normal) are considered to develop automated kidney segmentation and disease classification using CNN-based models. We utilized a balanced and limited number of images from a publicly available CT scan image dataset containing 12,446 images categorized into four classes (tumor, cyst, stone, and normal). The overall model development procedure has four major stages. The first stage, the data preprocessing stage, involves target mask preparation and data augmentation. The second stage involves training a U-Net segmentation model using the augmented dataset. In the third stage, CNN-based models, VGG-16, ResNet-50, and Inception-V3, were trained on segmented images generated using a segmentation model to classify KD. In this stage, batch normalization, dropout, regularization, and dense layers are added to optimize the models. Finally, the model is designed to provide possible expert-level treatment recommendations. The experiment results confirmed that using the threshold technique improved the performance of the U-Net model with an accuracy of 98%, and a Dice coefficient of 0.9. All the models were trained with the same amount of dataset and fixed hyperparameters; however, ResNet-50 outperformed with a testing accuracy of 97%.