Automated detection of kidney abnormalities using multi-feature fusion convolutional neural networks

摘要

Kidney abnormalities have a high incidence in people of all ages. The requisite manual examinations to detect these abnormalities are costly and time consuming. The rapid and accurate detection of kidney abnormalities has emerged as a focus of computer-aided medical research. Traditional methods work on images of kidneys and rely on hand-crafted features to identify abnormal symptoms as independent classes that lack generalization to different kidney diseases. In this study, an automated architecture to detect various kidney abnormalities is proposed that works on abdominal ultrasound images using convolutional neural networks. Our model consists of three components: the first selects appropriate ultrasound images of kidneys, the second is a detection model used to locate the area occupied by the kidney in the given image, and the third component using the multi-feature fusion neural network (Mf-Net) discriminates normal and abnormal kidneys, by determining whether there are abnormal symptoms in images. The detection model is combined with a weighted ensemble method to improve performance. A multi-feature fusion layer is also designed in the Mf-Net to extract distinctive features for multiple views of images. The three components work together to automatically recognize abnormalities associated with kidneys. A large dataset containing 3,722 abdominal ultrasound images with classification and localization annotations is established to train and evaluate the model. Experimental results show that the proposed ensemble detection model performs best with an average TPF of 98.0%, and the Mf-Net achieves an average classification accuracy of 94.67%. The obtained high classification and detection accuracy demonstrate the effectiveness of the proposed method for recognizing kidney abnormalities.