Regional Detection of Cervical Lesions Based on Self-attention Mechanism and Multi-scale Feature Enhancement

Abstract

Listen

Translate article

AbstractThe grade and affected area of cervical epithelial neoplasia (CIN) directly influence the probability of transformation into cervical cancer and the effect of early preventive treatment. This paper proposes a cervical lesion area detection network model (CT-YOLOv5) based on the improved YOLOv5s algorithm, accurately detecting the cervical disease area and disease level on cervical images collected by colposcopy. The transformers module was added at the end of the backbone feature extraction network to improve the ability to obtain different textual information of cervical images. The feature extraction part fuses the feature layer with each other layer through PANet. Using the detection strategy designed in this paper, a convolutional block attention module (CBAM) was added to the detection head to suppress irrelevant information, strengthen beneficial information and merge them into a new model, CT-YOLOv5. The CT-YOLOv5 was compared with commonly used detection models and models mentioned in related literature, such as SSD, YOLOv5, CLDNet, and the improved HLDNet. CT-YOLOv5 was also compared with networks using just transformers or CBAM. The precision, recall, and mAP scores of CT-YOLOv5 were 93.97%, 92.94%, and 92.8%, respectively, which outperform commonly used detection models and networks using just transformers or CBAM. CT-YOLOv5 also outperformed methods in relevant literature in both precision and recall. The computer-aided detection based on this method can help doctors accurately detect the diseased area and disease level on cervical images.KeywordsCervical lesion detectionDeep learningCBAMTransformers