UCSF

Medical imaging including magnetic resonance imaging (MRI) has become a major source of information for making clinical decisions, specifically based on static and dynamic shape variations of organs and moving structures respectively. Most conventional algorithms for shape extraction from images rely on deterministic modeling that offers no quantification for confidence of the extracted shapes. In this thesis a probabilistic approach for shape extraction is tested that provides a degree of confidence in extracted shapes. The degree of confidence create the important information in clinical decisions. This novel method, termed probabilistic Bayesian shape analysis introduced by Le.T and Schuff.N in [1], utilizes Curvelet transformation and Hidden Markov model simultaneously to detect probabilistic distributions of shape contours. In addition, this thesis, aims to effectively summarize probabilistic shape features in terms of information theoretic measures, such as the entropy (E) and statistical complexity (SC). The approach was initially demonstrated on well-defined hand gestures. To show the clinical potential of probabilistic shape feature extraction, the novel method was tested on tongue shape variations during pronunciation of vowels. Probabilistic shape analysis of tongue movement during vowel pronunciations mapped on MRI was performed on 5 subjects [1 woman, 4 men, age range 25-58], who speak English. MRI consisted of a Fast Low Angle Shot sequence [2]. All subjects were asked to sequentially pronounce the vowels [u:], [i:], [a:], [ae:], [e:], [o:], each for approximately 7 seconds, while having the fast MRI scan of the tongue. A probabilistic contour of tongue shape was extracted from each MRI frame (total = 210) and its features were summarized in terms of E and SC. Variations in E and SC as a function of vowel were then tested using univariate linear mixed effects regression. In addition, a multivariate linear mixed effects regression based on Monte Carlo sampling was used to test variations in E and SC simultaneously. First, the analysis showed that the novel method significantly captures variations in tongue shapes compared to repeated MRI variations. Second, the analysis demonstrated that entropy of the tongue extracted consistent shapes for the vowels [u:], [ae:], and [o:] across subjects, and for SC consistent features were found for [u:], [ae:], [a:] and [o:]. Moreover, using E and SC together capture consistent features for the vowels [a:], [e:], [i:] and [u:] across subjects. In conclusion, this thesis illustrated that the probabilistic shape model captures tongue shapes associated with vowel pronunciation. The new method is of potential interest to clinical studies of tongue disorders and function, including speech therapy, assessment of tongue surgery and assessment of functionality of the tongue. Beyond the tongue, the novel probabilistic approach has potentially wide clinical applications in almost any medical field that uses imaging. Examples of applications for the importance of contour extraction and shape detection include the delineation of almost any organ, such as the shape of kidneys and liver, the changing shapes of tendons in joints during movement, the tracking of change in structure of the beating heart and the shape of a tumor, specifically in the brain.