Medical Analytica Ltd, Flintshire, UK.
School of Computing, University of Buckingham, Buckingham, UK.
Ultrason Imaging. 2021 May;43(3):124-138. doi: 10.1177/0161734621998091. Epub 2021 Feb 25.
Significant successes in machine learning approaches to image analysis for various applications have energized strong interest in automated diagnostic support systems for medical images. The evolving in-depth understanding of the way carcinogenesis changes the texture of cellular networks of a mass/tumor has been informing such diagnostics systems with use of more suitable image texture features and their extraction methods. Several texture features have been recently applied in discriminating malignant and benign ovarian masses by analysing B-mode images from ultrasound scan of the ovary with different levels of performance. However, comparative performance evaluation of these reported features using common sets of clinically approved images is lacking. This paper presents an empirical evaluation of seven commonly used texture features (histograms, moments of histogram, local binary patterns [256-bin and 59-bin], histograms of oriented gradients, fractal dimensions, and Gabor filter), using a collection of 242 ultrasound scan images of ovarian masses of various pathological characteristics. The evaluation examines not only the effectiveness of classification schemes based on the individual texture features but also the effectiveness of various combinations of these schemes using the simple majority-rule decision level fusion. Trained support vector machine classifiers on the individual texture features without any specific pre-processing, achieve levels of accuracy between 75% and 85% where the seven moments and the 256-bin LBP are at the lower end while the Gabor filter is at the upper end. Combining the classification results of the top ( = 3, 5, 7) best performing features further improve the overall accuracy to a level between 86% and 90%. These evaluation results demonstrate that each of the investigated image-based texture features provides informative support in distinguishing benign or malignant ovarian masses.
在各种应用的图像分析中，机器学习方法取得了显著的成功，这激发了人们对医学图像自动诊断支持系统的浓厚兴趣。对肿瘤细胞网络纹理癌变变化方式的深入理解不断为这些诊断系统提供信息，使用更合适的图像纹理特征及其提取方法。最近，已经有几种纹理特征被应用于通过分析来自卵巢超声扫描的 B 模式图像来区分恶性和良性卵巢肿块，其性能水平也有所不同。然而，缺乏使用常见的临床认可图像集对这些报告特征进行比较性能评估的研究。本文对七种常用纹理特征（直方图、直方图矩、局部二值模式[256 -bin 和 59-bin]、方向梯度直方图、分形维数和 Gabor 滤波器）进行了实证评估，使用了一组 242 张具有不同病理特征的卵巢肿块超声扫描图像。评估不仅考察了基于单个纹理特征的分类方案的有效性，还考察了使用简单多数规则决策级融合的这些方案的各种组合的有效性。使用没有任何特定预处理的单个纹理特征训练支持向量机分类器，其准确性水平在 75%到 85%之间，其中七个矩和 256-bin LBP 处于较低端，而 Gabor 滤波器处于较高端。将排名前 （ = 3, 5, 7）的最佳表现特征的分类结果进行组合，可进一步将整体准确性提高到 86%到 90%之间。这些评估结果表明，所研究的基于图像的纹理特征中的每一种都为区分良性或恶性卵巢肿块提供了有价值的支持。
