The major problem addressed in this proposal is the development and evaluation of an automated noninvasive approach to discriminate different normal and pathological tissue types using machine learning algorithms;previous applications of machine learning have been based on features of the backscattered ultrasound that are essentially energy based. Our approach will be based on extracting features from images whose pixels are determined by the entropy contained in segments of the backscattered ultrasound. The unique attributes of entropy imaging suggest that the automated analysis we propose would be particularly robust for discrimination of deep tissues in a clinical environment.
All skilled clinical practitioners and interpreters of ultrasound studies realize that much information exists in recorded US images that is processed immediately by the visual cortex and is useful for qualitatively defining pathology, yet defies ready quantification by any robust algorithm. Traditional energy-based representations display grayscale intensities and speckle patterns that have been mapped parametrically into various tissue classification schemes that have yet to demonstrate organ or tissue specificity, although progress has been reported in distinguishing pathologies over the last 30 years. However, the fact that US signal processing and representation of backscatter data in terms of energy functions has not changed over the last 50 years suggests that alternative signal processing schemes may be indicated to represent the richness of the information contained within the backscattered data. To meet this challenge, we have been involved over the past 10 years in processing backscattered RF to create information images and in designing information sensitive approaches to classifying the data sets based on statistical analysis of these images These novel and user independent metrics utilize the entropy of windowed segments of radiofrequency (RF) backscatter signal from tis- sue, which represents a radical departure from grayscale or speckle metrics. In this approach the entropy of the backscattered segment is used to produce a pixel value in the tissue image. This processing strategy has proven to be sensitive to weak, sub-resolution sized changes in tissue.