Deep feature enhancement and Xgboost network for multi‐organ classification

Abstract

AbstractThe multi‐organ classification of medical images in disease examination and diagnosis is very valuable. However, some organs features of medical imaging are not obvious, and the organ of interest just occupy several pixels in the image. Therefore, the accuracy of multi‐organ classification of medical images using classic convolutional neural networks (CNN) cannot meet actual requirements. To address the above problems, a Deep Feature Enhancement and Xgboost (DFEX) network is proposed in this paper. The network enhances the discriminability of deep features through the dual‐core CNN feature extraction, the transfer module, and the multi‐level semantic feature fusion module. The dual‐core CNN feature extraction and transfer module extracts high‐level semantic features with detailed information by four aggregation units, to alleviate the problem of inconspicuous imaging features. The multi‐level semantic feature fusion module fuses the low‐level, middle‐level, and high‐level semantic features of the image to obtain more detailed features, which alleviates the problem that the target organ is easily affected by other organs. Finally, the enhanced features are fed to the Xgboost ensemble method for multi‐organ classification tasks of medical images. Our method is compared with 4 different classic networks on 553 cases. The experiment results show our method has better overall performance. Especially, for nasopharynx and vocal fold closing, the classification accuracies of our method achieve 84.39% and 88.00%. Compared with the best performers in classic networks, the accuracies have been increased by 3.58% and 15.62% respectively.