Cervical cell TCT image detection and segmentation based on multi-scale feature fusion
Meng Long, Guangming Liang, Yang Zheng, Zhineng Li, Jie Ming Zhong · 2021
Based on Mask R-CNN, this paper proposes a cervical image target detection and segmentation model based on multi-scale feature fusion. Mask R-CNN's backbone network-Feature Pyramid (FPN) uses top-down cross-layer connection, which causes the network to pay too much attention to the optimization of low-level features, resulting in a problem of reduced segmentation performance. Aiming at this problem, this paper uses NAS-FPN to determine the search space under the guiding ideology of classification network architecture search, realizes automatic cross-layer connection, and obtains the best feature fusion strategy. And use Soft NMS instead of NMS to reduce the mistaken deletion of bounding boxes and improve the accuracy of target detection. To The author produced his own TCT cell image data set (New Perdell's liquid-based cytological detection technique) under the guidance of medical laboratory personnel. After expanding the TCT data set, experiments show that the improved model in this paper can effectively improve the accuracy, recall and segmentation accuracy of epithelial cells, white blood cells, and fungi.