基于Cascade R-CNN的乳腺钼靶肿块检测算法研究

Research of Breast Molybdenum Target Mass Detection Algorithm Based on Cascade R-CNN

乳腺癌生物学特性复杂,恶性程度极高,位于女性恶性肿瘤发病率首位。乳腺钼靶肿块的X射线检查是早期确诊乳腺癌的重要方式。但乳腺钼靶肿块的检测尚处于早期阶段,现有的计算机辅助检测检测精度较低。针对这一问题,论文提出了一种基于Cascade R-CNN的乳腺钼靶肿块检测方法。实验使用南佛罗里达大学的乳房X光检查数据集,将乳腺钼靶肿块分为良性和恶性两类。通过在特征网络中加入注意力模块,提取了较为丰富的乳腺钼靶肿块特征。此外,论文提出了一种新的FPN网络FA-FPN,进一步提高了乳腺钼靶肿块病灶特征的提取,解决了深层网络在下采样中特征出现稀释的问题,提高了乳腺钼靶肿块的检测准确率。经实验验证,该模型在南佛罗里达大学的乳房X光检查数据集上的mAP值达到82.9%,在AP75下表现尤为突出。该方法在乳腺钼靶肿块的检测中具有良好的性能,可以提高乳腺钼靶肿块的检测精度,并在一定程度上避免了误检和漏检。

Breast cancer has complex biological characteristics and high malignancy,ranking the first place in the incidence rate of female malignant tumors.X ray examination of mammographic mass is an important way to diagnose breast cancer early.How⁃ever,the detection of breast molybdenum target mass is still in the early stage,and the existing computer-aided detection accuracy is low.To solve this problem,a breast molybdenum target mass detection method based on Cascade R-CNN is proposed in this pa⁃per.Using the breast X-ray data set of the University of South Florida,breast molybdenum target masses are divided into benign and malignant.By adding the attention module to the feature network,rich features of breast molybdenum target masses are extract⁃ed.In addition,this paper proposes a new FPN network FA-FPN,which further improves the extraction of lesion features of breast molybdenum target masses,and solves the problem that the biological characteristics of breast molybdenum target masses are com⁃plex and difficult to extract features.The experimental results show that the map value of the model on the breast X-ray data set of the University of South Florida reaches 82.9%,especially under AP75.This method has good performance in the detection of breast molybdenum target mass,can improve the detection accuracy of breast molybdenum target mass,and avoid false detection and missed detection to a certain extent.