We propose a general ultrasound (US) texture-analysis and machine-learning framework for detecting the presence of disease that is suitable for clinical application across clinicians, disease types, devices, and operators. Its stages are image selection, image filtering, ROI selection, feature parameterization, and classification. Each stage is modular and can be replaced with alternate methods. Thus, this framework is adaptable to a wide range of tasks.
Our two preliminary clinical targets are hepatic steatosis and adenomyosis diagnosis. For steatosis, we collected US images from 288 patients and their pathology-determined values of steatosis (%) from biopsies. Two radiologists independently reviewed all images and identified the region of interest (ROI) most representative of the hepatic echotexture for each patient.
To parameterize the images into comparable quantities, we filter the US images at multiple scales for various texture responses. For each response, we collect a histogram of pixel features within the ROI, and parameterize it as a Gaussian function using its mean, standard deviation, kurtosis, and skew to create a 36-feature vector. Our algorithm uses a support vector machine (SVM) for classification. Using a threshold of 10%, we achieved 72.81% overall accuracy, 76.18% sensitivity, and 65.96% specificity in identifying steatosis with leave-ten-out cross-validation (p<0.0001).
Extending this framework to adenomyosis, we identified 38 patients with MR-confirmed findings of adenomyosis and previous US studies and 50 controls. A single rater picked the best US-image and ROI for each case. Using the same processing pipeline, we obtained 76.14% accuracy, 86.00% sensitivity, and 63.16% specificity with leave-one-out cross-validation (p<0.0001).