Between Dec 19, 2024 and Jan 2, 2025, datasets can be submitted to DRUM but will not be processed until after the break. Staff will not be available to answer email during this period, and will not be able to provide DOIs until after Jan 2. If you are in need of a DOI during this period, consider Dryad or OpenICPSR. Submission responses to the UDC may also be delayed during this time.

Browsing by Subject "Medical Imaging"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    Machine Learning Methods with Emphasis on Cancerous Tissue Recognition
    (2018-08) Stanitsas, Panagiotis
    Today, vast and unwieldy data collections are regularly being generated and analyzed in hopes of supporting an ever-expanding range of challenging sensing applications. Modern inference schemes usually involve millions of parameters to learn complex real-world tasks, which creates the need for large annotated datasets for training. For several visual learning applications, collecting large amounts of annotated data is either challenging or very expensive; one such domain is medical image analysis. In this thesis, machine learning methods were devised with emphasis on Cancerous Tissue Recognition (CTR) applications. First, a lightweight active constrained clustering scheme was developed for the processing of image data which capitalizes on actively acquired pairwise constraints. The proposed methodology introduces the use of the Silhouette values, conventionally used for measuring clustering performance, in order to rank the degree of information content of the various samples. Second, an active selection framework that operates in tandem with Convolutional Neural Networks (CNNs) was constructed for CTR. In the presence of limited annotations, alternative (or sometimes complementary) venues were explored in an effort to restrain the high expenditure of collecting image annotations required by CNN-based schemes. Third, a Symmetric Positive Definite (SPD) image representation was derived for CTR, termed Covariance Kernel Descriptor (CKD) which consistently outperformed a large collection of popular image descriptors. Even though the CKD successfully describes the tissue architecture for small image regions, its performance decays when implemented on larger slide regions or whole tissue slides due to the larger variability that tissue exhibits at that level, since different types of tissue can be present as the regions grow (healthy, benign disease, malignant disease). Fourth, to leverage the recognition capability of the CKDs to larger slide regions, the Weakly Annotated Image Descriptor (WAID) was devised as the parameters of classifier decision boundaries in a multiple instance learning framework. Fifth, an Information Divergence and Dictionary Learning (IDDL) scheme for SPD matrices was developed for identifying appropriate geometries and similarities for SPD matrices and was successfully tested on a diverse set of recognition problems including activity, object, and texture recognition as well as CTR. Finally, a transition of IDDL to an unsupervised setup was developed, dubbed alpha-beta-KMeans, to address the problem of learning information divergences while clustering SPD matrices in the absence of labeled data.
  • Thumbnail Image
    Item
    Medical screening: medical imag[in]ing, the body, and the self.
    (2011-06) Cohen, Sara Jo
    This dissertation examines representations of medical imaging technologies in order to understand how sick artists (writers, painters, and filmmakers) use diagnostic technologies in their work to assert their subjectivity in the face of medicine's efforts to objectify them. It uses the phrase "medical imag[in]ing technology" to refer broadly to visual and aural aids that have assisted physicians and surgeons with their work, and focuses primarily on the X-ray, the microscope, and the telephone, which had a brief stint as a proto-ultrasound device at President James A. Garfield's deathbed. Its chapters engage a series of texts connected in their efforts to understand the body through medical imag[in]ing technologies: Thomas Mann's 1924 novel The Magic Mountain, Alice James's diary, Frida Kahlo's paintings, Sergio Leone's westerns, Norman Mailer's films and writings, and the narratives surrounding the James-Younger Gang's 1876 raid on Northfield, Minnesota, and the 1881 assassination of President Garfield. I posit that these nineteenth and twentieth century texts and their efforts to understand both illness and medicine through medical imaging technologies anticipate contemporary efforts to involve patients in their health care through online medical records, illness blogs, and illness-based social networking sites.
  • Thumbnail Image
    Item
    Modern Diagnostic and Therapeutic Ultrasound Systems: A Nonlinear Approach
    (2024-06) Sahoo, Abhishek
    The evolution of science, technology, and engineering has shaped every aspect of the world we live in. Medical science is no exception. Tremendous computing power, miniaturization of hardware electronics, advances in material science, etc. have improved the quality of life by bringing precise and affordable diagnostic imaging facilities that ensure the possibility of earlier treatment plans. Among the various traditional imaging modalities, ultrasound has been a familiar and popular choice due to its portability, safety, low cost, and non-invasiveness, to name a few. Traditionally, ultrasound is used for imaging favorable acoustic media such as soft tissues, the abdomen, blood vessels, etc. whereas Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) have been the standard methods for imaging bone, muscle, brain, etc. owing to their superior spatial resolution and the ability to display tiny anatomical details to diagnose medical conditions. Therefore, therapeutic applications of ultrasound such as neuromodulation and high-intensity focused ultrasound (HIFU) typically rely on CT and MR image guidance before locking a target and initiating surgical procedures. However, the advent of dual-mode ultrasound arrays (DMUA) brings ultrasound one step closer to being self-reliant by unlocking both diagnostic and therapeutic capabilities for a wide range of applications. The inherent registration and shared time base between the imaging and therapeutic sequences make DMUAs ideal for critical applications such as transcranial imaging and focused ultrasound delivery where precision is the key. For image-guided therapeutic ultrasound to succeed and become the standard medical practice in the future, diagnostic quality and image specificity enhancement of the DMUAs is an absolute necessity, which is going to be the main focus of this dissertation. A nonlinear approach to ultrasound image reconstruction is studied to demonstrate its advantages over the standard B-mode imaging method. Ultrasound imaging artifacts such as reverberation, multiple scattering, side-lobe and grating lobe artifacts, beamforming artifacts, ultrasound clutter, etc. are fundamental to the physics of acoustic propagation and the assumptions made by the image reconstruction algorithms. Their presence compromises the overall diagnostic value, and image specificity by reducing the tissue contrast. However, it is well known that the nonlinearly generated ultrasound echoes are purer and contain significantly fewer undesired artifacts than their fundamental counterpart. Hence, an ultrasound image reconstructed from nonlinear echo components is superior in imaging quality. Following the polynomial signal processing concepts, a quadratic Volterra filter is designed to extract second-order nonlinearities. The challenge of optimal signal separation between the fundamental and nonlinear energies faced in the previously reported work is overcome by a band-selective quadratic kernel synthesis approach. First, a data-driven least-squares quadratic Volterra kernel is estimated from the imaging data samples to separate the nonlinear echo components. The quadratic kernel is further decomposed using the singular value decomposition (SVD). Next, the individual singular modes are analyzed and grouped according to their bifrequency responses. The grouped singular modes are then filtered with the corresponding 2D Gaussian bandpass filters to remove the out-of-band noisy interactions followed by a coherent summation to broaden the effective bandwidth of the kernel. Multiple sets of quadratic images are reconstructed using each group of kernels which capture energies at different nonlinear frequency interactions. The efficiency of the band-selective quadratic Volterra kernel synthesis algorithm is demonstrated on several imaging datasets acquired using multiple ultrasound arrays and then measuring their spatial resolution, contrast-to-tissue ratio (CTR), artifacts suppression, etc. A quality assurance phantom image shows the recovery of the true size of the embedded structures, improved axial and lateral resolution, and sidelobe suppression with the nonlinearly reconstructed images. In another study, in an in-vivo imaging of a porcine kidney, the quadratically reconstructed images reveal the structure and boundary of the kidney that is otherwise invisible in the standard B-mode image. Thus, the Volterra based band-selective kernel synthesis approach is successfully applied to multiple datasets to reconstruct high-quality, artifacts-free images revealing the true structure and dimension of the anatomical structures. Volterra based nonlinear signal separation model is an excellent analytical tool that helps us connect the theory of nonlinear propagation of an ultrasound beam to a filter design problem for separating the echo components. Moving forward, an algorithm-driven filter design procedure is pursued to capture nonlinear frequency interactions across the entire bifrequency spectrum irrespective of their participating energy. We depart from the idea of a least-squares kernel estimation strategy where quadratic energy is usually biased by fundamental wave energy. Instead, multiple kernels are designed after identifying all possible bifrequencies of interest having a higher likelihood of quadratic interactions in the medium. This is done after a thorough spectral analysis, identification of fundamental and a few higher-order harmonics, and second-order pairing of all the frequency components to form bifrequency pairs. Then, two-dimensional bandpass filters around these bifrequencies of interest are synthesized. First, two-dimensional lowpass filters are designed as per filter specifications by extending one-dimensional FIR lowpass filter concepts and then the required amount of bifrequency shifts are incorporated by multiplying 2D weighting operators following the modulation property of Fourier transform. Finally, the 2D quadratic filters are applied to transskull and transcranial imaging datasets acquired from a series of experiments conducted on a couple of cadaveric human heads and the quadratic images are reconstructed. The performance of all the filters is compared by computing the bifrequency responses of the 2D filters, and the quality of the quadratic images. It is observed that among all the most likely candidates, only a few of the quadratic filters correspond to true quadratic energy interactions whereas the rest are noisy which is evident from the quality of the reconstructed images. Additionally, all the true filters capturing energy at different spectral interactions preserve different imaging features, giving us a comprehensive visualization of different anatomical structures for a better diagnostic assessment. In our case, this has led to precise image guidance during FUS delivery in the cortical region of the brain. In addition to diagnostic quality enhancement, nonlinear imaging methods have been successfully implemented in quantitative ultrasound applications such as displacement estimation with higher accuracy. The final phase of the research is driven in the direction of implementing the above quadratic filter in real-time on a field programmable gate array (FPGA). An eigenvalue decomposition based parallel cascade architecture is designed to approximate the quadratic filter with just a couple of FIR filters and squaring elements to reconstruct quadratic images. The proposed architecture runs with significantly fewer resources and at a higher speed compared to the traditional direct implementation of the filter. Additionally, this leads to a scalable solution where the resource and timing analyses are linear w.r.t the filter memory length. This successful demonstration opens the door for low-cost, low-power, real-time hardware implementation in wearable medical ultrasound applications.
  • Thumbnail Image
    Item
    Solving Inverse Problems in Medical Imaging Using Statistical Signal Processing and Deep Learning
    (2020-02) Hosseini, Seyed Amir Hossein
    Solving inverse problems remain an active research area in various fields to study the cause of a phenomenon by observing the effects. In particular, such efforts are well grounded in medical imaging applications where inverse problems naturally arise due to the imaging target being either inaccessible or invisible to human eyes. Non-regularity or ill-conditioning is a major challenge in such situations which is a direct consequence of limited observations/measurements being available. Medical imaging applications have classically incorporated domain-specific knowledge about the forward encoding operator to regularize the pertaining inverse problem. Deep learning has recently received high interest as an alternative approach for data-driven regularization. In principle, such regularization can be achieved by acquiring more measurements. However, sufficient measurements are usually unavailable due to practical considerations. For instance, in electromagnetic source imaging (ESI) using electroencephalography (EEG), number of recording electrodes can not be excessively increased due to subsequent distortions that electrodes contacts can cause in the medium. This limitation adversely affects the spatial resolution of EEG-based ESI. Another example is magnetic resonance imaging (MRI) where increasing imaging resolution trades off with noise properties, temporal resolution and scan duration. High-resolution MRI can become prohibitively long especially when a sequence of images need to be obtained. Thus, accelerated MRI techniques are required to reconstruct high-resolution images using less data. The present thesis contributes to solving regularization issues pertaining to inverse problems in medical imaging with a focus on high-resolution electromagnetic source imaging (ESI) using electroencephalography (EEG) and accelerated high resolution magnetic resonance imaging (MRI) with high fidelity. The ESI spatial resolution and the brain full coverage is shown to significantly improve when simultaneous scalp EEG and intracranial EEG recordings are jointly utilized. Additionally, a new ESI technique is proposed that incorporates forward modeling uncertainties into inverse problem formulations, considerably improving localization accuracy and recovering underlying source dynamics and extents when such uncertainties are properly determined and modeled. A novel deep learning-based reconstruction technique is proposed for accelerated MRI using a recurrent neural network architecture that offers desired noise properties compared to conventional techniques, particularly at higher acceleration rates. Furthermore, a new deep learning methodology is developed that unrolls inverse problem optimization algorithms into dense recurrent neural networks for improved reconstruction performance without requiring extra computational power or increasing reconstruction time.