傳統磁振成像重建依賴於 反向快速傅立葉變換（IFFT），雖然計算效率高（$O(N \log N)$），但必須在均勻的 笛卡兒格網上。然而，現代臨床需求——例如 鈉磁振成像 用於腫瘤偵測——需要 非笛卡兒軌跡 （螺旋/放射狀）來捕捉衰減極快的訊號。
1. 插值法與迭代解法的比較
由於螺旋軌跡的樣本無法對齊格網，我們無法直接應用IFFT。必須選擇使用 插值法 （將樣本透過 加權函數進行插值）或 迭代重建。後者由 Haldar 與 Liang提出，將重建視為線性求解問題：$$(F^H F + \lambda W^H W)\rho = F^H d$$
2. 計算架構的轉變
序列式中央處理器在臨床時間框架內無法應付迭代求解器的$O(N)$複雜度。透過轉向 GPU上的大量平行運算，我們可將每個 體素 體素對應至一個獨特執行緒，將原本深陷於嵌套複雜性的噩夢轉化為高效能吞吐量優化的核心運算。
main.py
TERMINALbash — 80x24
> Ready. Click "Run" to execute.
>
QUESTION 1
Why is the Fast Fourier Transform (FFT) insufficient for spiral MRI trajectories?
FFT is too slow for clinical use.
Spiral data is not sampled on a uniform rectangular grid.
Spiral signals do not have frequency components.
FFT cannot handle sodium signals.
✅ Correct!
FFT requires a uniform grid. Non-Cartesian data must be 'gridded' or solved iteratively.❌ Incorrect
The issue is geometry, not speed; FFT is actually faster but requires specific data alignment.QUESTION 2
What is the primary trade-off when moving from Cartesian to Non-Cartesian scanning?
Higher SNR vs Lower Resolution.
Faster Acquisition vs Higher Reconstruction Complexity.
Lower Cost vs Higher Patient Motion.
Better SNR vs Slower Data Collection.
✅ Correct!
Non-Cartesian paths collect data faster and are robust to motion but require massive GPU power to reconstruct.❌ Incorrect
Actually, Non-Cartesian often improves motion robustness and sampling efficiency.QUESTION 3
In the parallel FHd kernel, what is mapped to individual CUDA threads?
K-space samples (m)
Image voxels (n)
The entire spiral trajectory
The apodization function
✅ Correct!
Mapping voxels to threads allows each thread to accumulate its own sum, avoiding race conditions.❌ Incorrect
Mapping samples to threads would require atomic additions because multiple samples contribute to one voxel.QUESTION 4
Which MRI application specifically benefits from spiral trajectories due to fast signal decay?
Standard T1 Brain Scans
Sodium MRI for tumor detection
Knee Ligament Scans
Static Bone Density Mapping
✅ Correct!
Sodium signals decay rapidly; spirals capture the k-space center immediately before the signal vanishes.❌ Incorrect
Sodium protons vanish much faster than hydrogen, necessitating the speed of spiral paths.QUESTION 5
What role does the 'Apodization Function' play in Gridding?
It speeds up the CPU clock.
It handles the interpolation of irregular samples to a grid.
It measures the signal-to-noise ratio.
It performs the final matrix inversion.
✅ Correct!
It acts as a kernel function to spread non-grid sample data onto a uniform grid for FFT processing.❌ Incorrect
Apodization is a mathematical interpolation step, not a hardware or SNR measurement tool.Clinical Case Study: The Sodium Map (Figure 8.2)
Optimizing Reconstruction for Early Tumor Detection
A surgical team needs a high-resolution sodium map of a patient's brain (Figure 8.2) to identify metabolic changes in a suspected tumor. Sodium signals decay in milliseconds, requiring a spiral trajectory. Traditional reconstruction takes 2 hours on a CPU, which is too slow for the surgical window.
Q
1. Why does a standard IFFT fail to produce Figure 8.2 when data is collected via a spiral?
Solution:
The Fourier transform assumes a rectilinear grid. Spiral data points are distributed non-uniformly across k-space, meaning there is no direct row-column relationship required for the IFFT algorithm.
The Fourier transform assumes a rectilinear grid. Spiral data points are distributed non-uniformly across k-space, meaning there is no direct row-column relationship required for the IFFT algorithm.
Q
2. Explain how Massive Parallelism (GPU) reduces the 'Clinical Bottleneck' in this scenario.
Solution:
By using a voxel-to-thread mapping, the GPU can compute the contribution of every k-space sample to every voxel simultaneously. This transforms the 2-hour sequential linear solver into a near-real-time parallel process.
By using a voxel-to-thread mapping, the GPU can compute the contribution of every k-space sample to every voxel simultaneously. This transforms the 2-hour sequential linear solver into a near-real-time parallel process.