Abstract
Quantitative imaging in MRI usually involves acquisition and reconstruction of a series of images at multi-echo time points, which possibly requires more scan time and specific reconstruction technique compared to conventional qualitative imaging. In this work, we focus on optimizing the acquisition and reconstruction process of multi-echo gradient echo pulse sequence for quantitative susceptibility mapping as one important quantitative imaging method in MRI. A multi-echo sampling pattern optimization block extended from LOUPE-ST is proposed to optimize the k-space sampling patterns along echoes. Besides, a recurrent temporal feature fusion block is proposed and inserted into a backbone deep ADMM network to capture the signal evolution along echo time during reconstruction. Experiments show that both blocks help improve multi-echo image reconstruction performance.
Introduction
Quantitative Imaging (QI) in MRI is a powerful tool that measures physical tissue properties, such as magnetic susceptibility (QSM) or relaxation times (T1/T2), rather than just producing a qualitative image. However, QI is notoriously slow because it requires a multi-echo pulse sequence to track signal evolution over time. To accelerate this process, researchers use "under-sampling," skipping parts of the raw k-space data. This presents a dual challenge: determining which data points are most critical to acquire and developing reconstruction techniques that can accurately "fill in the blanks" by exploiting the complex relationships between echoes.
This paper introduces a unified deep learning framework that simultaneously optimizes both acquisition and reconstruction. By combining a Sampling Pattern Optimization (SPO) block with a Temporal Feature Fusion (TFF) recurrent module inside a Deep ADMM network, the researchers significantly improved the speed and quality of multi-echo reconstruction, producing clearer images with fewer artifacts than traditional methods.
Methods
The proposed framework consists of three main components designed to leverage the temporal and spatial redundancies inherent in multi-echo MRI data.
Deep ADMM Backbone
The system uses the Alternating Direction Method of Multipliers (ADMM), a classic optimization algorithm, "unrolled" into a deep neural network. This allows the model to learn optimal reconstruction parameters directly from data, rather than relying on hand-crafted mathematical assumptions.
Sampling Pattern Optimization (SPO)
Based on the LOUPE-ST framework, the SPO block learns the optimal under-sampling mask for k-space. Instead of using a generic pattern, the network learns which data points are most critical for multi-echo reconstruction under a fixed budget (e.g., a 23% sampling rate). The "Multi SPO" approach allows for learned patterns that vary across echoes, capturing more information than static masks.
Temporal Feature Fusion (TFF)
The TFF block is a recurrent neural network (RNN) module that captures signal evolution over time. Since multi-echo signals follow a specific physical decay model, the TFF module passes information from one echo to the next, treating the sequence like a video to help denoise and stabilize the final images.
Results
The study utilized a dataset from 7 human subjects scanned on a 3T GE system with a 10-echo MEGRE sequence. The data was divided into 600 slices for training, 200 for validation, and 600 for testing, with fully sampled k-space data serving as the ground truth.
Quantitative and Qualitative Performance
The proposed "Deep ADMM + TFF + Multi SPO" model achieved a PSNR of 43.75 and an SSIM of 0.9894, representing a substantial jump in fidelity over the baseline Deep ADMM (PSNR 40.95, SSIM 0.9820). These improvements were statistically significant ($p < 0.05$).
Visually, the method eliminated heavy artifacts that plagued traditional methods like Locally Low Rank (LLR). As shown in Figure 2, the proposed approach better preserved fine "veil structures" in the brain, which are critical for accurate clinical diagnosis in QSM. The learned masks from the Multi SPO block significantly outperformed manually designed variable-density patterns, proving that joint optimization of acquisition and reconstruction is key.
Discussion and Conclusion
This work demonstrates that optimizing the sampling mask alongside the reconstruction network yields superior results compared to optimizing either in isolation. By exploiting temporal redundancy through the TFF block and using AI to find efficient scanning patterns, the framework achieves high-quality quantitative imaging from significantly under-sampled data. Future work will focus on testing the model on larger, more diverse patient populations and exploring its generalizability to other quantitative mapping techniques like T1 or T2 mapping.
Summary updated