REAF: ROI Extraction and Adaptive Fusion for Breast Cancer Diagnosis in Ultrasound Images
Ziyang Zhang, Jia Wei Lim, Yuanxun Zheng, Bozhao Chen, Dongxin Chen, Yi Lin · 2023
Breast cancer leads to the top cause of death among women worldwide. Early detection of breast cancer is the most effective way to reduce the risk of death and increase the likelihood of survival. Breast ultrasound imaging is an essential test for early detection but its accurate evaluation is difficult due to the limited number of experienced sonographers. So the computer-aided system for breast cancer diagnosis on ultrasound images is vitally necessary. Recent breast cancer diagnosis methods are mostly developed by improving or transferring visual deep models to breast cancer diagnosis. These methods ignored the characteristics of breast cancer on ultrasound images. Some methods fuse the pre-defined region of interest (ROI) to adapt the characteristics of ultrasound images but cannot diagnose those samples whose ROIs are not defined. In this paper, we propose a novel framework, ROI Extraction and Adaptive Fusion (REAF), for breast cancer diagnosis on ultrasound images. To make the framework be applied to the samples without pre-defined ROIs as well as incorporate the characteristics of breast cancer on ultrasound images, the proposed REAF firstly extracts the ROI of breast tumor areas as a confidence degree map. Then we propose an adaptive fusion method to fuse the extracted ROI into the downstream diagnostic model. In addition, we propose a pair augmentation method to increase the variety of samples for enhancing the generalization of the proposed REAF. The experimental results demonstrate that the REAF surpasses state-of-the-art visual classification models and breast cancer methods. For the evaluation of the REAF, the AUC reaches 95.52%, the accuracy reaches 88.97%, the specificity reaches 92.86%, the precision reaches 89.43%, the recall reaches 86.82% and the f1-score reaches 88.05%.