Improvement of Resolution of Ultrasound images

    This work has been supported by NSF under grant MIP-9553227, the US ARMY under grant DAMD17-94-J-4362, and the Whitaker Foundation

    We consider the problem of improving the resolution of ultrasound (US) images, so that ultrasound can be used for early breast cancer (BC) detection. X-ray mammography is the primary tool for early BC detection, with the exception of young patients or patients with dense breasts. Due to its poor resolution, US is currently used only as a complement to X-ray, to differentiate cystic from solid masses. The enhancement of the US resolution is of paramount importance because US offers many advantages (e.g., non-ionizing, relatively inexpensive and uses widely available, portable, equipment), which become even more important if one takes into account the disagreement regarding the use of ionizing X-rays for screening, or routine exams. The breast, however, is a particularly difficult organ for ultrasound to examine. During an ultrasonic investigation a pressure field is emitted into the tissue, interacts with it and is subsequently reflected back to the transducer. The received field forms the US image. Small sound speed variations between fat and other soft tissues distort the sound beam and change its path slightly, leading to poor images. Finite bandwidth of the transducer, spreading and aberrations of the ultrasound beam are mainly responsible for the loss of resolution. The US image can be viewed as a distorted version of the actual tissue response, the distortion being unknown. Enhancing resolution is equivalent to reconstructing the tissue response. We approach this as a blind deconvolution problem, employing multiple US images of the same tissue obtained by varying the transducer characteristics (frequency, aperture shape or size), or the transducer spatial arrangement. This allows us to identify in-vivo the spatially varying distortion (both lateral as well as axial), and subsequently cancel it in order to restore the tissue response. The long term goal of this effort is to establish US as a safe, accurate and economically attractive alternative to X-ray mammography for early breast cancer detection. We investigate various signatures of the tissue response and develop robust criteria for detection of small targets inside the tissue.
    Since ultrasound is widely used in nondestructive testing of materials, improving its resolution will also be beneficial in areas such as flaw detection in metals and material characterization.

    This research is funded by the National Science Foundation, the Whitaker Foundation, and the US army Medical Research Command.

    Collaborators
    Dr. John M. Reid , Biomedical Engineering and Science Inst., Drexel University
    Dr. Flemming Forsberg, Radiology Department, Thomas Jefferson University Hospital, Philadelphia PA.
    Dr. Barry Goldberg, Radiology Department, Thomas Jefferson University Hospital, Philadelphia PA.
     

    Some results on distrotion estimation and subsequent image deconvolution are shown below.

    ULTRASOUND RF ECHO MODELING AND TISSUE CHARACTERIZATION

    Supported by the National Institute of Health under grant grant CA52823

    Ultrasound is a widely used medical imaging technique because of its low cost, relative safety, and versatility. Since biological tissues are composed of characteristic structures whose ultrasonic properties often change due to diseases, the ultrasound RF echo contains information that can be used to study the underlying tissue. The goal is to model and process the ultrasound RF echo in order to extract tissue characterization features that are observer-independent.

    In this project, we propose a new model for the radio-frequency (RF) ultrasound echo, namely the shot noise process with narrow-band power-law filter function. This model can be justified by considering the tissue as a collection of point scatterers embedded in a uniform medium and assuming a power-law decay instead of exponential decay for attenuation. The model is characterized by the exponent n of the power-law filter and Poisson rate l. Based on this model, the in-phase and quadrature components of the echo can be shown to exhibit 1/f b -type spectral behavior with b = 2(1-n). The envelope also exhibits this type of spectral behavior but with a different exponent and furthermore if the power-law exponent n is equal to 0.5 the envelope follows the well-known Rayleigh statistic. Although the shot noise model has been used in the past for modeling the RF echo, this is the first time that a power-law impulse response filter is used and the resulting power-law spectral behavior of the RF echo is investigated. The theoretical derivations were validated via simulations, while the validity of the proposed model was tested based on clinical ultrasound images of the breast.

    The spectral exponents in the proposed model are associated to  tissue attenuation, whereas l is associated to the number of scatterers. Since both properties change due to disease, the estimates of these parameters are  natural candidates for tissue signatures. We have proposed algorithms for estimating b and l from the data, and investigated the potential significance of the model parameters in characterizing breast tissue. We conducted experiments based on clinical ultrasound images of the breast, provided to us by our collaborators at Thomas Jefferson University Hopsittal in Philadelphia PA. The model parameters were first estimated based on a database of 100 clinical and Receiver Operating Characteristic (ROC) analysis was subsequently applied to quantify their ability to differentiate between tumorous and non-tumorous tissue. High detection probabilities  at low false-alarm rates (97.1% at 4.2% for the spectral exponent of the envelope, and 84.3% at 17.1% for the spectral exponent of the in-phase component), and respectively 97% and 93%   ROC area values were obtained, which suggest that these parameters can e used for tissue characterization. The performance of the parameter  l (ROC area of 89%) was not as good as that of the spectral exponents.

    Although the results on differentiating between normal and abnormal tissue are very good, the ability of model parameters in differentiating between malignant and benign tumor regions is moderate. ROC areas of 72.5%, 70%, and 68% were obtained respectively for the exponent of the envelope, the in-phase component and l. Our results were based on 56 images, 28 of which contained benign tumors and rest contained malignant tumors. We are currently investigating the amount of independent information carried by b and l  and the potential advantages of combining the two parameters to obtain a better tissue signature.

Features