BETTER THAN REAL: COMPLEX-VALUED NEURAL NETS FOR MRI FINGERPRINTING【记录】

题目： 优于真实:用于mri指纹识别的复杂值神经网络

Fig. 1. MRI fingerprinting is an inverse mapping problem that
infers the tissue parameters from MRI signals. MRI simulator
turns a ground-truth (T1, T2, B0) parameter tuple into an
observed MRI temporal signal. The inverse mapping, by either
nearest neighbor search or a neural network, solves for the
tissue parameters given the simulated signal. At test time,
the MRI signal will arrive from the scanner rather than the
simulator. Example complex-valued signals are shown for
cerebral spinal fluid (CSF) and white matter (WM) parameters.

特点：

1. 虚数为0的情况下就是第三张图上的黄色线，就是实数域上的ReLU激活函数

2. 复值从正实轴旋转到负实轴的相位上的过程中，虚部从非零值逐渐缩放为0

3. 保留了复值信号的相位分量，但衰减了复值信号的幅度【引用自22年的综述Complex-Valued Neural Networks: A Comprehensive Survey】

激活函数表达式

网络结构：

实验结果：

能看出复数网络在重建T1 T2上很准确， 近邻搜索方法在重建B0上更好

近邻搜索法(字典)随着待测参数的增加，计算效率骤减

误差图像(右边)越清晰(越亮)就证明误差越大，从上图看出，字典的方法对于预测B0很在行。