光电工程, 2019, 46 (1): 180368, 网络出版: 2019-01-18
基于超声RF信号的乳腺肿瘤分级检测方法
The grade classification algorithm of breast tumor based on ultrasound RF signals
计算机辅助诊断 超声RF信号 支持向量机 Shearlet变换 CAD ultrasound RF signal support vector machine Shearlet transformation
摘要
为解决超声乳腺肿瘤分级检测问题, 从超声射频(RF)信号的角度提出了一种有效的乳腺肿瘤分级检测方法。首先, 采用Shearlet变换提取乳腺超声RF信号的多尺度、多方向特征; 其次, 考虑Shearlet特征的高维冗余性, 采用多尺度方向二值模式(MDBP)对其进行编码, 在不损失特征信息的条件下降低特征维度; 最后, 依据医生阅片经验以及不同分级乳腺肿瘤的特征差异性, 设计出适合乳腺病变分级检测的层级二叉树SVM分类器(CBT-SVM)。在928个乳腺肿瘤患者的超声RF信号上进行验证, 大量结果表明, 提出方法可以有效实现3级、4A级~4C级、5级乳腺肿瘤的分级检测, 准确度、敏感度、特异度、PPV、NPV以及MCC分别达到89.29%、75.62%、94.54%、97%、98.3%和81.01%。
Abstract
A novel efficient method based on the ultrasound radio frequency (RF) signals is proposed to distinguish the breast tumors grades. First, we utilize the multi-scale geometric characteristic of Shearlet transformation to extract the multi-scale and multi-directional features of ultrasound RF signal, and then reduce the high-dimensional Shearlet features by multi-scale directional binary pattern which can effectively preserve the sufficient discriminated information. At last, we draw on the feature difference between different grades of breast tumors to design a cascade binary tree SVM classifier which not only overcome the problem of sample quantity disequilibrium but also conform to the subjective diagnosis rule of sonographer. Extensive experiments on 928 breast ultrasound RF signals collected from the hospital demonstrate the effectiveness of the new proposed method and its precision, sensitivity, specificity, PPV, NPV and MCC are 89.29%, 75.62%, 94.54%, 97%, 98.3% and 81.01%, respectively.
童莹, 严郁. 基于超声RF信号的乳腺肿瘤分级检测方法[J]. 光电工程, 2019, 46(1): 180368. Tong Ying, Yan Yu. The grade classification algorithm of breast tumor based on ultrasound RF signals[J]. Opto-Electronic Engineering, 2019, 46(1): 180368.