A Breast Ultrasound Tumor Detection Framework Using Convolutional Neural Networks
Hongguang Yang, Xudong Wang, Jiyong Tan, Gen Liu, Xi Hong Sun, Yuanwei Li · 2022
Accurate and efficient breast cancer screening is of great significance to women's health. In order to solve the severe challenges in mass breast screening, such as poor ultrasound image quality, differences in the age and geographical distribution of the population, we proposed a detection framework based on convolution neural networks for tumor detection and tracking in ultrasound video. Firstly, some data pre-processing and tricks are adjust to improve YOLOv4 for making it more suitable for tumor detection task. Secondly, Kernelized Correlation Filters (KCF) tracking algorithm as post-processing is used to track and fuse all the detection bounding boxes. In this way, all the detection results can be aggregated to form a smaller number of tumor sequences, and some false positives can also be filtered out. The proposed method was evaluated on 251 cases with tumors. It obtains a promising result with sensitivity 97.62% and 12.3 false positives per case. Experimental results demonstrate that our method has better performance on tumor detection for ultrasound videos from mass breast screening.