Breast lesion detection using an anchor-free network from ultrasound images with segmentation-based enhancement

Abstract

The survival rate of breast cancer patients is closely related to the pathological stage of cancer. The earlier the pathological stage, the higher the survival rate. Breast ultrasound is a commonly used breast cancer screening or diagnosis method, with simple operation, no ionizing radiation, and real-time imaging. However, ultrasound also has the disadvantages of high noise, strong artifacts, low contrast between tissue structures, which affect the effective screening of breast cancer. Therefore, we propose a deep learning based breast ultrasound detection system to assist doctors in the diagnosis of breast cancer. The system implements the automatic localization of breast cancer lesions and the diagnosis of benign and malignant lesions. The method consists of two steps: 1. Contrast enhancement of breast ultrasound images using segmentation-based enhancement methods. 2. An anchor-free network was used to detect and classify breast lesions. Our proposed method achieves a mean average precision (mAP) of 0.902 on the datasets used in our experiment. In detecting benign and malignant tumors, precision is 0.917 and 0.888, and recall is 0.980 and 0.963, respectively. Our proposed method outperforms other image enhancement methods and an anchor-based detection method. We propose a breast ultrasound image detection system for breast cancer detection. The system can locate and diagnose benign and malignant breast lesions. The test results on single dataset and mixed dataset show that the proposed method has good performance.

BUS‐BRA: A breast ultrasound dataset for assessing computer‐aided diagnosis systems

“…Lesion detection aims to locate the bounding box that encloses the ROI containing a lesion. Recent approaches have employed generic CNN‐based object detectors (e.g., Fast R‐CNN, YOLO, and SSD) 63 and specific‐purpose architectures 64 . On the other hand, lesion segmentation outlines the lesion shape, where generic semantic segmentation models (e.g., SegNet, UNet, and DeepLab) 33 have been used, and specific‐purpose approaches have been developed 65 …”

“…Recent approaches have employed generic CNN-based object detectors (e.g., Fast R-CNN, YOLO, and SSD) 63 and specific-purpose architectures. 64 On the other hand, lesion segmentation outlines the lesion shape, where generic semantic segmentation models (e.g., SegNet, UNet, and DeepLab) 33 have been used, and specificpurpose approaches have been developed. 65 Lesion classification methods usually distinguish between pathology classes interpreted as conducting a follow-up imaging study if the lesion is classified as benign or prescribing a biopsy if the lesion is classified as malignant.…”

Semi‐supervised learning for ABUS tumor detection using deep learning method

“…In Zhang et al [15], the 3D ResNet was added after the tumor detection of YOLOv5 for false-positive reduction, which was trained by a two-stage manner. Different from the anchor based detector, Wang et al [16] employed the anchor-free network FCOS for tumor detection. Besides, an image enhancement method was designed to improve the image contrast.…”

“…Different from the anchor based detector, Wang et al. [16] employed the anchor‐free network FCOS for tumor detection. Besides, an image enhancement method was designed to improve the image contrast.…”

A Robust Quality Improvement Framework of Breast Ultrasound Images for Detection of Tumours

“…Frequency domain de-speckling is based on wavelet transform that converts the continuous-time signal into different frequency components (wavelets), essentially converting the speckles into additive noise and performing despeckling in the frequency domain. Yu Wang et al [3] presented an image enhancement algorithm to elevate the visual quality of the image using Contrast limited adaptive histograms equalization (CLAHE) to enhance the Breast ultrasound images (BUSI) and then if any corners of the image are left to enhance it is achieved with the help of the Anisotropic Diffusion. Further, the classification is done through multiple algorithms such as U-Net with an accuracy of 71.2% and Recurrent Residual convolutional Neural Based on u-net (R2U-Net) with an accuracy of 71.1%.…”