UCSF

We propose an adaptable framework for analyzing ultrasound (US) images quantitatively to provide computer-aided diagnosis using machine learning. Our preliminary clinical targets are hepatic steatosis, adenomyosis, and craniosynostosis. For steatosis and adenomyosis, we collected US studies from 288 and 88 patients, respectively, as well as their biopsy or magnetic resonanceconfirmed diagnosis. Radiologists identified a region of interest (ROI) on each image. We filtered the US images for various texture responses and use the pixel intensity distribution within each ROI as feature parameterizations. Our craniosynostosis dataset consisted of 22 CT-confirmed cases and 22 age-matched controls. One physician manually measured the vectors from the center of the skull to the outer cortex at every 10 deg for each image and we used the principal directions as shape features for parameterization. These parameters and the known diagnosis were used to train classifiers. Testing with cross-validation, we obtained 72.74% accuracy and 0.71 area under receiver operating characteristics curve for steatosis ([Formula: see text]), 77.27% and 0.77 for adenomyosis ([Formula: see text]), and 88.63% and 0.89 for craniosynostosis ([Formula: see text]). Our framework is able to detect a variety of diseases with high accuracy. We hope to include it as a routinely available support system in the clinic.