Multi-source data fusion for recognition of cervical precancerous lesions

Cervical cancer is a major threat to women's health. Fusion of patients' multi-source data is helpful to improve the diagnostic accuracy of cervical precancerous lesions. In this end, we propose a novel convolutional neural network framework that fuses the features of the two-source images (acetic acid and iodine test images) to identify cervical precancerous lesions. In our framework, we first leverage two ResNet to extract spatial features to obtain feature maps. Then, We have developed a feature selection module based on attention mechanism to integrate the related information in the feature maps. To obtain the global feature description, we leverage factorized bilinear pooling technique to fuse the feature maps of different data sources. We make use of a cervical dataset of 2,800 clinical data to train and evaluate our proposed method. The result shows that the fusion of information from multi-source data is effective. Also, the proposed method has better performance than the related approaches.