Transfer learning in deep neural network-based receiver coil sensitivity map estimation.

INTRODUCTION: The success of parallel Magnetic Resonance Imaging algorithms like SENSitivity Encoding (SENSE) depends on an accurate estimation of the receiver coil sensitivity maps. Deep learning-based receiver coil sensitivity map estimation depends upon the size of training dataset and generalization capabilities of the trained neural network. When there is a mismatch between the training and testing datasets, retraining of the neural networks is required from a scratch which is costly and time consuming. MATERIALS AND METHODS: A transfer learning approach, i.e., end-to-end fine-tuning is proposed to address the data scarcity and generalization problems of deep learning-based receiver coil sensitivity map estimation. First, generalization capabilities of a pre-trained U-Net (initially trained on 1.5T receiver coil sensitivity maps) are thoroughly assessed for 3T receiver coil sensitivity map estimation. Later, end-to-end fine-tuning is performed on the pre-trained U-Net to estimate the 3T receiver coil sensitivity maps. RESULT AND CONCLUSION: Peak Signal-to-Noise Ratio, Root Mean Square Error and central line profiles (of the SENSE reconstructed images) show a successful SENSE reconstruction by utilizing the receiver coil sensitivity maps estimated by the proposed method.

Accelerating GRAPPA reconstruction using SoC design for real-time cardiac MRI.

Real-time cardiac MRI is a rapidly developing area of research that has the potential to improve the diagnosis and treatment of cardiovascular diseases. However, the acquisition of high-quality real-time cardiac MR (CMR) images is challenging as it requires a high frame rate and temporal resolution. To overcome this challenge, there have been recent efforts on several approaches including hardware-based improvements and image reconstruction techniques such as compressed sensing and parallel MRI. The use of parallel MRI techniques such as GRAPPA (Generalized Autocalibrating Partial Parallel Acquisition) is a promising approach for improving the temporal resolution of MRI and expanding its applications in clinical practice. However, the GRAPPA algorithm involves a significant amount of computation, particularly for high acceleration factors and large datasets. This can result in long reconstruction times, which can limit the ability to achieve real-time imaging or high frame rates. One solution to this challenge is to use specialized hardware i.e. field-programmable gate arrays (FPGAs). In this work, a novel 32-bit floating-point FPGA-based GRAPPA accelerator is proposed with an aim to reconstruct high-quality cardiac MR images at higher frame rates, making it well suited for real-time clinical applications. The proposed FPGA-based accelerator consists of custom-designed data processing units named as dedicated computational engines (DCEs) that allow for a continuous flow of data between the calibration and synthesis stages of GRAPPA reconstruction process. This greatly increases the throughput and reduces the latency of the overall proposed system. Moreover, a high-speed memory module (DDR4-SDRAM) is integrated with the proposed architecture to store the multi-coil MR data. An on-chip quad-core ARM Cortex-A53 processor is used to manage access control information required for data transfer between the DCEs and DDR4-SDRAM. The proposed accelerator is implemented on Xilinx Zynq UltraScale + MPSoC using high-level synthesis (HLS) and hardware descriptive language (HDL) with an aim to explore the trade-offs between the reconstruction time, resource utilization and design effort. Several experiments have been performed using in-vivo cardiac datasets i.e. 18-receiver coil and 30-receiver coil to evaluate the performance of the proposed accelerator. A comparison is performed with the contemporary CPU and GPU-based GRAPPA reconstruction methods in terms of reconstruction time, frames-per-second and reconstruction accuracy (RMSE and SNR). The results show that the proposed accelerator achieves speed-up factors up to 121× and 9× as compared to the contemporary CPU-based and GPU-based GRAPPA reconstruction methods, respectively. Moreover, it has been demonstrated that the proposed accelerator can achieve reconstruction rates of up to ∼27 frames-per-second while maintaining the visual quality of the reconstructed images.

GPU accelerated Cartesian GRAPPA reconstruction using CUDA.

BACKGROUND AND OBJECTIVE: GRAPPA (Generalized Auto-calibrating Partially Parallel Acquisition) is an advanced parallel MRI reconstruction method (pMRI) that enables under-sampled data acquisition with multiple receiver coils to reduce the MRI scan time and reconstructs artifact free image from the acquired under-sampled data. However, the reduction in MRI scan time comes at the expense of long reconstruction time. It is because the GRAPPA reconstruction time shows exponential growth with increasing number of receiver coils. Consequently, the conventional CPU platforms may not adhere to the requirements of fast data processing for MR image reconstruction. METHODS: Graphics Processing Units (GPUs) have recently emerged as a viable commodity hardware to reduce the reconstruction time of pMRI methods. This paper presents a novel GPU based implementation of GRAPPA using custom built CUDA kernels, to meet the rising demands of fast MRI processing. The proposed framework exploits intrinsic parallelism in the calibration and synthesis phases of GRAPPA reconstruction process, aiming to achieve high speed MR image reconstruction for various GRAPPA configuration settings using different number of receiver coils, auto-calibration signals (ACS), sizes of GRAPPA kernel and acceleration factors. In-vivo experiments (using 8, 12 and 30 receiver coils) are performed to compare the performance of the proposed GPU accelerated GRAPPA with the CPU based GRAPPA extensions and GPU counterpart. RESULTS: The results indicate that the proposed method achieves up to ≈47.8× , ≈17× and ≈3.8× speed up gains over multicore CPU (single thread), multicore CPU (8 thread) and Gadgetron (GPU based GRAPPA) respectively, without compromising the reconstruction accuracy. CONCLUSIONS: The proposed method reduces the GRAPPA reconstruction time by employing the calibration phase (GRAPPA weights estimation) and synthesis phase (interpolation) on GPU. Our study shows that the proposed GPU based parallel framework for GRAPPA reconstruction provides a solution for high-speed image reconstruction while maintaining the quality of the reconstructed images.

Transfer learning in deep neural network based under-sampled MR image reconstruction.

In Magnetic Resonance Imaging (MRI), the success of deep learning-based under-sampled MR image reconstruction depends on: (i) size of the training dataset, (ii) generalization capabilities of the trained neural network. Whenever there is a mismatch between the training and testing data, there is a need to retrain the neural network from scratch with thousands of MR images obtained using the same protocol. This may not be possible in MRI as it is costly and time consuming to acquire data. In this research, a transfer learning approach i.e. end-to-end fine tuning is proposed for U-Net to address the data scarcity and generalization problems of deep learning-based MR image reconstruction. First the generalization capabilities of a pre-trained U-Net (initially trained on the human brain images of 1.5 T scanner) are assessed for: (a) MR images acquired from MRI scanners of different magnetic field strengths, (b) MR images of different anatomies and (c) MR images under-sampled by different acceleration factors. Later, end-to-end fine tuning of the pre-trained U-Net is proposed for the reconstruction of the above-mentioned MR images (i.e. (a), (b) and (c)). The results show successful reconstructions obtained from the proposed method as reflected by the Structural SIMilarity index, Root Mean Square Error, Peak Signal-to-Noise Ratio and central line profile of the reconstructed images.

FPGA-based hardware accelerator for SENSE (a parallel MR image reconstruction method).

SENSE (Sensitivity Encoding) is a parallel MRI (pMRI) technique that allows accelerated data acquisition using multiple receiver coils and reconstructs the artifact-free images from the acquired under-sampled data. However, an increasing number of receiver coils has raised the computational demands of pMRI techniques to an extent where the reconstruction time on general purpose computers becomes impractically long for real-time MRI. Field Programmable Gate Arrays (FPGAs) have recently emerged as a viable hardware platform for accelerating pMRI algorithms (e.g. SENSE). However, recent efforts to accelerate SENSE using FPGAs have been focused on a fixed number of receiver coils (L=8) and acceleration factor (Af=2). This paper presents a novel 32-bit floating-point FPGA-based hardware accelerator for SENSE (HW-ACC-SENSE); having an ability to work in coordination with an on-chip ARM processor performing reconstructions for different values of L and Af. Moreover, the proposed design provides flexibility to integrate multiple units of HW-ACC-SENSE with an on-chip ARM processor, for low-latency image reconstruction. The VIVADO High-Level-Synthesis (HLS) tool has been used to design and implement the HW-ACC-SENSE on the Xilinx FPGA development board (ZCU102). A series of experiments has been performed on in-vivo datasets acquired using 8, 12 and 30 receiver coil elements. The performance of the proposed architecture is compared with the single thread and multi-thread CPU-based implementations of SENSE. The results show that the proposed design withstands the reconstruction quality of the SENSE algorithm while demonstrating a maximum speed-gain up to 298× over the CPU counterparts in our experiments.

QR-decomposition based SENSE reconstruction using parallel architecture.

Magnetic Resonance Imaging (MRI) is a powerful medical imaging technique that provides essential clinical information about the human body. One major limitation of MRI is its long scan time. Implementation of advance MRI algorithms on a parallel architecture (to exploit inherent parallelism) has a great potential to reduce the scan time. Sensitivity Encoding (SENSE) is a Parallel Magnetic Resonance Imaging (pMRI) algorithm that utilizes receiver coil sensitivities to reconstruct MR images from the acquired under-sampled k-space data. At the heart of SENSE lies inversion of a rectangular encoding matrix. This work presents a novel implementation of GPU based SENSE algorithm, which employs QR decomposition for the inversion of the rectangular encoding matrix. For a fair comparison, the performance of the proposed GPU based SENSE reconstruction is evaluated against single and multicore CPU using openMP. Several experiments against various acceleration factors (AFs) are performed using multichannel (8, 12 and 30) phantom and in-vivo human head and cardiac datasets. Experimental results show that GPU significantly reduces the computation time of SENSE reconstruction as compared to multi-core CPU (approximately 12x speedup) and single-core CPU (approximately 53x speedup) without any degradation in the quality of the reconstructed images.

Iterative Schemes to Solve Low-Dimensional Calibration Equations in Parallel MR Image Reconstruction with GRAPPA.

GRAPPA (Generalized Autocalibrating Partially Parallel Acquisition) is a widely used parallel MRI reconstruction technique. The processing of data from multichannel receiver coils may increase the storage and computational requirements of GRAPPA reconstruction. Random projection on GRAPPA (RP-GRAPPA) uses random projection (RP) method to overcome the computational overheads of solving large linear equations in the calibration phase of GRAPPA, saving reconstruction time. However, RP-GRAPPA compromises the reconstruction accuracy in case of large reductions in the dimensions of calibration equations. In this paper, we present the implementation of GRAPPA reconstruction method using potential iterative solvers to estimate the reconstruction coefficients from the randomly projected calibration equations. Experimental results show that the proposed methods withstand the reconstruction accuracy (visually and quantitatively) against large reductions in the dimension of linear equations, when compared with RP-GRAPPA reconstruction. Particularly, the proposed method using conjugate gradient for least squares (CGLS) demonstrates more savings in the computational time of GRAPPA, without significant loss in the reconstruction accuracy, when compared with RP-GRAPPA. It is also demonstrated that the proposed method using CGLS complements the channel compression method for reducing the computational complexities associated with higher channel count, thereby resulting in additional memory savings and speedup.