Projects

Fast and Accurate Quantitative Susceptibility Mapping

Quantitative Susceptibility Mapping (QSM) enables accurate quantification of myelin, iron, and calcium content in the brain by estimating the tissue magnetic susceptibility distributions from Magnetic Resonance (MR) phase measurements. To improve the quality of the QSM reconstruction, model resolution deconvolution was introduced. This is a two-step approach, wherein the model resolution matrix is utilized to estimate a closer approximation of the accurate susceptibility map. The performance of this deconvolution was evaluated using an approximated inverse, iterative framework, and a sparsity prior, and has been shown to improve the image quality over existing methods. Furthermore, to improve the speed of reconstruction in the iterative QSM formulation, the vector polynomial extrapolation method was used in the context of the state-of-the-art iterative method. To improve the quality of deep-learning-based QSM reconstructions, a Schatten p-norm-based regularizer was enforced in the QSM formulation. This has been shown to yield better performance in terms of quantitative metrics and image quality compared with state-of-the-art deep learning approaches. To address the inherent bias towards the training data distribution in the existing deep learning-based QSM methods or lack of generalizability, a test time correction approach was proposed without adapting the model weights. In this method, the susceptibility map prediction obtained from the deep learning model is refined iteratively to ensure consistency with the measured local field, imposing model-based data consistency. The performance of the proposed approach was also evaluated using susceptibility map predictions obtained from state-of-the-art deep learning-based QSM reconstruction approaches such as QSMnet, learned proximal convolutional neural network (LPCNN), and fidelity imposed network edit (FINE). 

Median Nerve Segmentation in Ultrasound Videos

Carpal Tunnel Syndrome (CTS) is the most common peripheral  neuropathy, which affects the thumb, index finger, middle finger, and lateral side of the ring finger. Typically, this arises due to the increased pressure within the enclosed carpal tunnel that compresses the median nerve and is characterized by tingling and numbness of the affected hand. Ultrasound allows high-resolution imaging of the median nerve and the carpal tunnel and is a commonly used technique for the diagnosis of Carpal Tunnel Syndrome. Typically, the median nerve cross-sectional area (CSA) is used as a parameter to diagnose CTS. In this project, we utilize an efficient deep-learning model to automatically segment the median nerve at the wrist and compute the CSA. The commonly accepted rule of thumb being CTS is diagnosed if CSA is greater than 12 mm2 our model automatically computes the CSA and enables an automated diagnosis. This model has shown an accuracy of approximately 92%. Further, the clinical evaluation of the developed model is performed on a larger cohort of unseen data consisting of subjects with and without CTS.

A Deep Learning based Back Project Filter Method for Limited Angle Computed Tomography

Computed tomography (CT) is an efficient imaging tool that plays an important role in medical diagnosis, industrial inspection, and security checks. Accurate reconstruction of a CT image requires the test object to be scanned at least under consecutive 180° or 180°+fan angles for parallel-beam or fan-beam geometries. To reduce harmful radiation doses, limited angle (LA) acquisition with a decreased number of projection views is a preferable choice in medical applications. However, with such acquisitions, conventional CT reconstruction approaches, such as filtered back projection (FBP) and iterative reconstruction, exhibit limited angle artifacts. To overcome the difficulties associated with the abovementioned conventional CT reconstruction approaches, we propose a back-project-filter-based reconstruction wherein the deconvolution operation is performed using a convolutional neural network (CNN). The workflow of the proposed approach is explained as follows. Initially, the limited-angle sinogram was back projected to obtain a blurred CT image. The back-projected CT image was then fed into a deep symmetric encoder-decoder architecture (UNet) to obtain a CT image with reduced blurring. This image was forward-projected to obtain a sinogram with an extended number of views (e.g., views corresponding to 180°+60°, i.e., 481 views were used throughout this approach). Subsequently, a simultaneous iterative reconstruction technique (SIRT) was performed on the extended sinogram to obtain the final reconstruction. 

An alternating minimization approach for implementation of the proximal continuation-based regularizer

The estimation of iteration-dependent parameters for continuation is outlined as an optimization problem by the inclusion of an extra prior to the compressed sensing (CS) magnetic resonance imaging (MRI) cost function. An alternating minimization approach is formulated to solve the optimization problem, in which the first sub-problem updates the sparse domain solution and the second sub-problem updates the regularization parameter. Whereas the solution to the first sub-problem is obtained using the proximal splitting methods, the second one is computed using root-finding techniques such as the modified Regula-Falsi method that can ensure faster convergence to the required solution. The reconstruction error obtained using the proposed approach converges to that computed using the mean-squared error (MSE) optimal constant regularization (benchmark solution), independent of the initially chosen regularization parameter.

Sparsity promoting adaptive regularization for accelerated convergence

An iteration-dependent parameter selection strategy is developed for achieving simultaneous improvements in both speed and steady-state errors in sparsity-based formulations. For this, it is required to use a large value of the threshold at the start of the iterations that are decreased in the succeeding iterations, until the target regularization is achieved. To achieve automatic parameter selection, a parameter update rule is derived that utilizes the non-negativity of discrepancy level (absolute difference between l1-norms of consistency and sparse approximation errors). Application of adaptive regularization combined with over-relaxation to parallel MRI shows significant improvements in both speed and image quality.

Frequency-dependent regularization for autocalibrating parallel MRI

A frequency-dependent form of regularized reconstruction is developed consisting of two phases. In the first phase of cross-over determination, noise perturbations are introduced into the calibration data. In each perturbation step, the calibration matrix is constructed in a manner to satisfy the generalized discrepancy principle (GDP) condition, requiring the residual error norm to exceed the norm of perturbation error. The perturbation step in which both residual and perturbation error norms are nearly equal is referred to as the cross-over point. Imposition of an upper limit on the extent of regularization at each missing k-space location is achieved by matching the cross-over error bound to perturbation of k-space values resulting from reconstruction using Tikhonov filters with varying increments of regularization parameter from that of a reference filter.