Breast tumor detection in ultrasound images with anatomical prior knowledge

Abstract

Breast tumor detection is an important step in the procedure of computer-aided diagnosis. In clinical practice, computer-aided diagnosis system not only processes lesion images but also processes normal images without lesions. However, normal images are often overlooked. In this study, we additionally collected numerous normal images to evaluate object detection algorithms. We found that similarities between tumors and hypoechoic regions have led to false positive lesions, and the frequency of false positive lesions in normal images is higher than it in lesion images. To address this issue, we incorporate anatomical prior knowledge of breast tumors to propose a novel breast tumor detection method. Our method consists of a preprocessing method and a novel breast tumor detection network. The preprocessing method automatically extracts breast regions as anatomical constraints and utilizes channel fusion to combine images of breast regions with original images. The proposed breast tumor detection network is based on programmable gradient information and large-kernel convolution. The programmable gradient information is applied by an auxiliary branch which provides more comprehensive gradient information for backpropagation, while large-kernel convolution expands the receptive field of neurons. As a result, our method achieves the best false positive lesion rate of 3.30% and gets a reduction by at least 5.67% over other compared algorithms in normal images, with the best precision of 90.91%, sensitivity of 88.57%, f1-score of 89.75%, and mean average precision of 93.07% in lesion images. Experimental results suggest promising application potential.