基于阈值图像分割结合深度学习的血细胞识别算法研究

目的:探讨传统图像处理的阈值图像分割与深度学习相结合的血细胞识别算法,用于血细胞涂片的自动识别和分类。方法:使用全局阈值分割从血细胞涂片中提取血细胞并单独保存,对分割后的细胞图像进行人工标记和分类,以降低后续处理的硬件要求。标记后的图像深度学习训练基于GoogLeNet预训练模型,生成可用于血细胞涂片图像自动识别的深度学习模型。使用已训练模型对测试集进行评价,并生成混淆矩阵和受试者工作特征(ROC)曲线下面积(AUC)值。结果:这种预处理被证明可以提高深度学习模型的训练以及后续使用模型识别的速度10倍以上。使用在线血细胞涂片图像数据集Raabin-WBC Data,模型训练的准确率达到93.06%,均取得了良好的结果。结论:基于阈值图像分割与深度学习相结合的血细胞识别算法,可极大提高血细胞识别分类的效率,保证血液相关疾病诊断的准确度。

基于超声射频信号的乳腺肿瘤分级算法研究

为解决乳腺肿瘤超声的定量分级问题,从超声射频信号的角度提出了一种乳腺肿瘤分级的评价算法。以乳腺影像报告和数据系统(Breast imaging reporting and data system,BI-RADS)作为分级依据,将提取的超声射频(Radio frequency,RF)信号进行图像重建、图像分割并获取乳腺肿瘤感兴趣区(Region of interest,ROI)及其特征参数:熵和标准差。量化分析特征参数与病灶良恶性分级之间的关系,实现了对乳腺肿瘤的3级、4级、5级的分级,分类成功率达到84.9%。研究结果表明,超声射频信号对辅助临床诊断具有重要意义,熵和标准差可以有效地实现乳腺肿瘤超声分级。

Breast Lump Recognition Algorithms Based on Ultrasound Radio-Frequency Signals

A method for evaluating the benign and malignant breast tumors based on radio?frequency(RF)data was explored by extracting the characteristic parameters of breast ultrasound RF signals.The breast biopsy data were used as the reference data for judging the lump benign or malignant.The extracted ultrasound RF data were reconstructed and segmented by computer aided method to obtain the breast tumor region of interest(ROI)and its characteristic parameters(entropy and standard deviation).The characteristic parameters were statistically analyzed to evaluate the relationship between characteristic parameters and benign or malignant breast tumors.The results indicate the entropy and standard deviation of normal region is much higher than that of lump region,which shows that the standard deviation and entropy characteristic parameters of ultrasonic RF signals are meaningful in the diagnosis of breast tumors.The proposed method provides a new direction for computer?aided diagnosis of benign and malignant breast tumors.