Research on Breast Cancer Detection Methods Based on ODMV-MulDyHead-YOLO

Abstract

Breast cancer has the highest incidence rate among women worldwide. Early screening and timely treatment can effectively control breast cancer progression and significantly reduce mortality. In this paper, we proposed a method of breast cancer detection based on full-dimensional dynamic convolution and multiple attention mechanism to solve the problems of missing detection and low detection accuracy caused by breast tumor occlusion by breast muscle, poor contrast between tumor and surrounding glandular tissue and indistinct features of small tumor. First, a target detection head based on an attention mechanism is proposed to improve lump detection sensitivity in complex backgrounds. Second, a lightweight backbone network, MobileNetV2-ODConv, is used to improve the ability of the network to extract features of lumps in breast X-ray images while reducing the number of parameters and floating-point operations. Finally, a loss function based on the minimum point distance is used to improve target localization accuracy and convergence speed. The experimental results show that the improved model achieves higher accuracy, sensitivity, and detection speed than previous models, and reduces false positive and false negative rates. In addition, two-core convolution is used instead of the normal 2D convolution in the C3 module. The experimental results show that: the number of parameters of the improved model is reduced by about 47%, and the recall rate and average detection accuracy of the improved model in DDSM dataset, MIAS dataset and INBreast dataset are increased by 8.2% and 6.7%, 3.7% and 2.2%, 1.3% and 2.4% respectively.

Keywords

• Breast Lump Detection
• Deep Learning
• Attention Mechanism
• MobileNetV2
• Dynamic Head
• X-ray Image