CTIFI: Clinical-experience-guided three-vision images features integration for diagnosis of cervical lesions

• Aiming at the high similarity of cervical lesions, a more effective feature extraction network SE-DenseNet is used to suppress the invalid features and enhance the effective features. • In view of the fact that many previous studies neglected the correlation between three-vision images in clinic, according to the guidance of clinical experience, a new cervical lesion network, CTIFI, was designed. • As for the limitations of clinical application, CTIFI classify the four lesion grades of Normal, LSIL, HSIL and Cancer, which can effectively help clinicians make diagnosis. At present, the research on diagnosis of cervical lesions based on deep learning mostly uses single-vision images or full-mixed images, ignoring the correlation among the three-vision images in the clinic, so that the effect is not good and the help to clinicians is extremely limited. Therefore, according to the guidance of clinical experience, this paper proposes a novel method of three-vision images features integration (CTIFI) for the classification and diagnosis of cervical lesions by simultaneously performing feature learning on three-vision images of the same patient. Firstly, SE-DenseNet is used to extract the features from three-vision cervical images. During this process, the invalid features are suppressed while the network is concentrated to important features. Then, the three-vision images features are integrated to effectively improve the performance of lesion classification. Under the same study conditions, this method was compared with other methods and clinicians. The results show that the accuracy (ACC) and the area under the curve (AUC) of this method were 71% and 0.876, which are superior to the average level of other methods and clinicians. Therefore, it can help clinicians make diagnosis, reduce misdiagnosis and missed diagnosis, so as to improve work efficiency.