Multi-model data fusion for cervical precancerous lesions detection

Cervical cancer ranks fourth among women in the world, seriously affecting women's health. Colposcopy is one of the main methods to find the location of cervical precancerous lesions, but it requires a lot of doctor resources and requires doctors to have rich experience.In the state of art, the lesion information can usually be recognized by deep learning on colposcopy images stained with acetic acid. However, the colposcopy image contains two stained images(called acetic images and iodine images), use only one of them will prone to false positive regions. Using deep learning to fuse patient acetic images and iodine images helps to improve the detection speed and accuracy of cervical precancerous lesions. To this end, we propose a novel detection framework. In our framework, first, we use target detection algorithm to detect the location of cervical cancer that may exist, secondly, we propose a new method to determine the location of cervical cancer to exclude false positive areas. In order to ensure that the target detection algorithm better learns the characteristics of cervical cancer lesions, we improved the Faster-RCNN algorithm by using methods such as average pooling and retention of decimals generated in the image calculation process. The experimental results show the performance of the proposed HLDNet, which achieves the average accuracy of 86%(sensitivity 82% and specificity 90%) in the test set including 200 real patients.