To provide a real-time SENSE (Sensitivity Encoding) reconstruction on FPGA (Field Programmable Gate Array) with the coil sensitivity maps estimated using (1) pre-scan method (2) Eigen-value approach

In parallel magnetic resonance imaging (MRI) scan time is reduced by acquiring fewer lines in k-space, which reduces the field-of-view (FOV) producing aliased (under-sampled) images. SENSE¹, is a widely used algorithm in clinical MRI scanners to reconstruct the unfolded image from under-sampled data. The importance of speed in parallel MRI real-time reconstruction generates the requirement to develop application specific hardware for SENSE reconstruction. This paper presents an FPGA implementation for SENSE reconstruction right on the receiver coil system without the need to transfer data to the server (workstation). In this paper, SENSE reconstruction is performed using two different coil estimation methods (pre-scan¹ and Eigen value decomposition method²). In the proposed architecture, 16-bit fixed point arithmetic binary notation is used to represent the real and imaginary parts of the complex MRI data. All the hardware modules such as complex multiplier, complex matrix multiplier, pseudo inverse and square root are designed according to the algorithmic needs to increase the throughput of the architecture.

The proposed architecture is tested on human head and phantom datasets acquired using GE MR450, 1.5T MRI scanner at St Mary’s Hospital London using Fast Spin Echo sequence with the following scan parameters: Slice thickness $$$3mm$$$, Matrix size $$$256×256$$$, Flip Angle $$$90°$$$, TR $$$520ms$$$, TE $$$15ms$$$ and FOV $$$55mm$$$. The proposed architecture uses coil sensitivity profiles obtained from the pre-scan method¹ and Eigen value decomposition method² (E-maps) for SENSE reconstruction. E-maps method is a better choice in situations where motion artefacts are likely to affect the data because it does not require a pre-scan and uses auto-calibration lines in k-space for sensitivity profile estimation². However, pre-scan method is computationally less complex to implement but requires additional low resolution images before the actual scan and is also sensitive to motion artefacts.

The coil sensitivity matrices and folded images are stored in the memory of the FPGA (Virtex-6) and the proposed SENSE architecture is used to reconstruct the un-folded image. Finally, the output is transferred from FPGA to MATLAB, using serial communication, for validation of the proposed system. Artefact power³ (AP) and SNR maps⁴ are used to quantify the quality of the reconstructed images.

Eight channel receiver coil datasets (phantom and human head) with an acceleration factor of 2 are used for the validation of the proposed system. The dimensions of the aliased images and sensitivity maps used in this paper are $$$128×256×8$$$ and $$$256×256×8$$$, respectively. A comparison between the reconstructed images computed using coil sensitivity maps obtained from both the methods are illustrated in Figure 1. Figure 2 shows the SNR maps of the reconstructed images. Table 1 provides a comparison of mean SNR and AP of the reconstructed images using pre-scan and E-maps sensitivity maps. The results confirm that the proposed design produces high quality reconstructed images from the under-sampled data.

The results show that the proposed architecture produces artefact-free images. One advantage of the Eigen-value method to estimate sensitivity maps is that it does not require a pre-scan which makes it suitable for the cases where motion artefacts are considerable. However, the pre-scan method has slightly better quality than the images reconstructed with E-maps in terms of AP and mean SNR. Smaller values of AP $$$(<9×10^{-4})$$$ and good mean SNR $$$(>29dB)$$$ ensure the quality of the reconstructed images produced by the proposed system.

The proposed system is capable for run-time SENSE reconstruction using both sensitivity maps estimation method and consumes merely $$$145.64 μs$$$ to produce reconstructed image $$$(256×256)$$$. Furthermore, the proposed system can be embedded on the receiver coil system leaving no need to transfer all the raw data to the memory banks of the MRI workstation, thereby, reducing the coaxial noise. The on-chip reconstruction not only provides efficient memory usage by only sending the reconstructed image but also reduces the data transmission cost.

The proposed FPGA based SENSE architecture has the potential to compute accurate images from under-sampled data using the receiver coil sensitivities (pre-scan sensitivity and E-maps sensitivity). Efficient memory usage and less transmission cost with remarkable mean SNR and AP are the significant features of the proposed work. The proposed architecture may be useful for the modern MRI systems.