Multi-view weighted feature fusion with wavelet transform and CNN for enhanced CT image recognition

Abstract

Reducing noise in CT images and extracting key features are crucial for improving the accuracy of medical diagnoses, but it remains a challenging problem due to the complex characteristics of CT images and the limitations of existing methods. It is worth noting that multiple views can provide a richer representation of information compared to a single view, and the unique advantages of the wavelet transform in feature analysis. In this study, a novel Multi-View Weighted Feature Fusion algorithm called MVWF is proposed to address the challenge of enhancing CT image recognition utilizing wavelet transform and convolutional neural networks. In the proposed approach, the wavelet transform is employed to extract both detailed and primary features of CT images from two views, including high frequency and low frequency. To mitigate information loss, the source domain is also considered as a view within the multi-view structure. Furthermore, AlexNet is deployed to extract deeper features from the multi-view structure. Additionally, the MVWF algorithm introduces a balance factor to account for both specific information and global information in CT images. To accentuate significant multi-view features and reduce feature dimensionality, random forest is used to assess feature importance followed by weighted fusion. Finally, CT image recognition is accomplished using the SVM classifier. The performance of the MVWF algorithm has been compared with classical multi-view algorithms and common single-view methods on COVID-CT and SARS-COV-2 datasets. The experimental results indicate that an average improvement of 6.8% in CT image recognition accuracy can be achieved by utilizing the proposed algorithm. Particularly, the MVF algorithm and MVWF algorithm have attained AUC values of 0.9972 and 0.9982, respectively, under the SARS-COV-2 dataset, demonstrating outstanding recognition performance. The proposed algorithms can capture more robust and comprehensive high-quality feature representation by considering feature correlations across views and feature importance based on Multi-view.