IR-CNN: Inception residual network for detecting kidney abnormalities from CT images

In this study, we propose an efficient architecture “IR-CNN” based on the Inception residual network for the detection of three major kidney diseases, tumor, kidney stone and cyst, using CT images. We customized the top layer of InceptionResNetV2 and further added global average pooling (GAP), batch normalization (BN), dropout and dense layers wi th swish activation functions to extract robust features, avoid vanishing gradient problems and achieve better accuracy in detecting kidney disease. The proposed IR-CNN model was trained and tested on a publicly available kidney CT dataset with 4000 images using different optimizers (Adam, SGD, and RMSprop). Experimental results show that IR-CNN achieves 99.38%, 94.63%, 97.38% using Adam, SGD and RMSprop optimizers, respectively. In addition, IR-CNN with Adam optimizer achieved better performance with only 5 misclassifications out of 800 test images and performed better than existing methods i n diagnosing kidney disease. The superior results of our IR-CNN architecture can help urologists diagnose kidney disease, thereby reducing human error.
Source: Network Modeling Analysis in Health Informatics and Bioinformatics - Category: Bioinformatics Source Type: research