Muhtadi Sabiq, Razzaque Rezwana R, Chowdhury Ahmad, Garra Brian S, Kaisar Alam S
University of North Carolina at Chapel Hill and North Carolina State University, Joint Department of Biomedical Engineering, Chapel Hill, North Carolina, United States.
Washington University in St. Louis, McKelvey School of Engineering, St. Louis, Missouri, United States.
J Med Imaging (Bellingham). 2023 Feb;10(Suppl 2):S22410. doi: 10.1117/1.JMI.10.S2.S22410. Epub 2023 Jun 22.
We evaluate texture quantified from ultrasound Nakagami parametric images for non-invasive characterization of breast tumors, as Nakagami images can more faithfully represent intrinsic tumor characteristics than standard B-mode images.
Parametric images were formed using sliding windows applied to ultrasound envelope data. To analyze the trade-off between spatial resolution and stability of estimated Nakagami parameters for texture quantification, two different window sizes were used for image formation: (i) the standard square window with sides equal to three times the pulse length of incident ultrasound, and (ii) a smaller square window with sides equal to exactly the pulse length. Texture was quantified from two different regions of interest (ROIs) consisting of the tumor core and a 5 mm surrounding margin. A total of 186 texture features were analyzed for each ROI, and feature selection was used to identify the most relevant feature sets for breast tumor characterization.
Texture quantified from parametric images formed using the two different windows did not outperform each other by a significant margin. However, when the mean pixel value within the tumor region of the parametric images was incorporated with the texture features, texture quantified from the tumor core and surrounding margin of images formed using the standard square window thoroughly outperformed other considerations for breast lesion characterization. The highest performing set of texture and mean value features yielded a significant AUC of 0.94, along with sensitivity of 90.38% and specificity of 89.58%.
We establish that texture quantified from ultrasound Nakagami parametric images are diagnostically relevant and may be used to characterize breast lesions effectively.
我们评估从超声中值参数图像量化得到的纹理，用于乳腺肿瘤的无创特征分析，因为中值图像比标准B模式图像能更忠实地反映肿瘤的内在特征。
使用应用于超声包络数据的滑动窗口形成参数图像。为了分析用于纹理量化的估计中值参数在空间分辨率和稳定性之间的权衡，使用两种不同的窗口大小进行图像形成：（i）边长等于入射超声脉冲长度三倍的标准方形窗口，以及（ii）边长恰好等于脉冲长度的较小方形窗口。从由肿瘤核心和5毫米周围边缘组成的两个不同感兴趣区域（ROI）量化纹理。对每个ROI总共分析186个纹理特征，并使用特征选择来识别用于乳腺肿瘤特征分析的最相关特征集。
使用两种不同窗口形成的参数图像量化得到的纹理在性能上没有显著差异。然而，当将参数图像肿瘤区域内的平均像素值与纹理特征结合时，从使用标准方形窗口形成的图像的肿瘤核心和周围边缘量化得到的纹理在乳腺病变特征分析方面明显优于其他考量。表现最佳的纹理和平均值特征集产生了显著的曲线下面积（AUC）为0.94，灵敏度为90.38%，特异性为89.58%。
我们证实，从超声中值参数图像量化得到的纹理具有诊断相关性，可有效用于乳腺病变的特征分析。
