Fan Beam Iterative Reconstruction
This example demonstrates 2D fan beam iterative reconstruction using the differentiable FanProjectorFunction and FanBackprojectorFunction from diffct.
Overview
Fan beam iterative reconstruction extends the optimization approach to the more realistic fan beam geometry. This example shows how to:
Formulate fan beam CT reconstruction as an optimization problem
Handle geometric complexities of divergent ray geometry
Apply gradient-based optimization with fan beam operators
Monitor convergence and reconstruction quality
Mathematical Background
Fan Beam Iterative Formulation
The fan beam reconstruction problem is formulated as:
where \(A_{\text{fan}}\) is the fan beam forward projection operator accounting for divergent ray geometry.
Fan Beam Forward Model
The fan beam projection operator maps 2D image \(f(x,y)\) to sinogram \(p(\beta, u)\):
where integration follows the ray from point source to detector element \(u\) at source angle \(\beta\).
Gradient Computation
The gradient involves the fan beam backprojection operator (adjoint):
where \(A_{\text{fan}}^T\) is the fan beam backprojection operator.
Geometric Considerations
Fan beam geometry introduces complexities compared to parallel beam:
Ray Divergence: Non-parallel rays affect sampling density and conditioning
Magnification Effects: Variable magnification across the field of view
Non-uniform Resolution: Spatial resolution varies with distance from rotation center
Geometric Distortion: Requires careful handling of coordinate transformations
Implementation Steps
Geometry Setup: Configure fan beam parameters (SID, SDD)
Problem Formulation: Define parameterized image and fan beam forward model
Loss Computation: Calculate L2 distance using FanProjectorFunction
Gradient Computation: Use automatic differentiation through fan beam operators
Optimization: Apply Adam optimizer with appropriate learning rate
Convergence Monitoring: Track reconstruction quality and loss evolution
Model Architecture
The fan beam reconstruction model consists of:
Parameterized Image: Learnable 2D tensor representing the unknown image
Fan Beam Forward Model: FanProjectorFunction with geometric parameters
Loss Function: Mean squared error between predicted and measured sinograms
Convergence Characteristics
Fan beam reconstruction typically exhibits:
Initial Convergence (0-100 iterations): Rapid loss decrease, basic structure
Detail Refinement (100-500 iterations): Fine features develop, slower progress
Final Convergence (500+ iterations): Minimal improvement, convergence plateau
Challenges and Solutions
Conditioning: Fan beam system matrix may have different conditioning properties
Geometric Artifacts: Proper weighting and filtering help reduce artifacts
Parameter Tuning: Learning rate may need adjustment for optimal convergence
Memory Usage: Similar to parallel beam but with additional geometric computations
1import math
2import torch
3import numpy as np
4import matplotlib.pyplot as plt
5import torch.nn as nn
6import torch.optim as optim
7from diffct.differentiable import FanProjectorFunction
8
9
10def shepp_logan_2d(Nx, Ny):
11 Nx = int(Nx)
12 Ny = int(Ny)
13 phantom = np.zeros((Ny, Nx), dtype=np.float32)
14 ellipses = [
15 (0.0, 0.0, 0.69, 0.92, 0, 1.0),
16 (0.0, -0.0184, 0.6624, 0.8740, 0, -0.8),
17 (0.22, 0.0, 0.11, 0.31, -18.0, -0.8),
18 (-0.22, 0.0, 0.16, 0.41, 18.0, -0.8),
19 (0.0, 0.35, 0.21, 0.25, 0, 0.7),
20 ]
21 cx = (Nx - 1)*0.5
22 cy = (Ny - 1)*0.5
23 for ix in range(Nx):
24 for iy in range(Ny):
25 xnorm = (ix - cx)/(Nx/2)
26 ynorm = (iy - cy)/(Ny/2)
27 val = 0.0
28 for (x0, y0, a, b, angdeg, ampl) in ellipses:
29 th = np.deg2rad(angdeg)
30 xprime = (xnorm - x0)*np.cos(th) + (ynorm - y0)*np.sin(th)
31 yprime = -(xnorm - x0)*np.sin(th) + (ynorm - y0)*np.cos(th)
32 if xprime*xprime/(a*a) + yprime*yprime/(b*b) <= 1.0:
33 val += ampl
34 phantom[iy, ix] = val
35 phantom = np.clip(phantom, 0.0, 1.0)
36 return phantom
37
38class IterativeRecoModel(nn.Module):
39 def __init__(self, volume_shape, angles,
40 num_detectors, detector_spacing,
41 sdd, sid, voxel_spacing):
42
43 super().__init__()
44 self.reco = nn.Parameter(torch.zeros(volume_shape))
45 self.angles = angles
46 self.num_detectors = num_detectors
47 self.detector_spacing = detector_spacing
48 self.sdd = sdd
49 self.sid = sid
50 self.relu = nn.ReLU() # non negative constraint
51 self.voxel_spacing = voxel_spacing
52
53 def forward(self, x):
54 updated_reco = x + self.reco
55 current_sino = FanProjectorFunction.apply(updated_reco, self.angles,
56 self.num_detectors, self.detector_spacing,
57 self.sdd, self.sid, self.voxel_spacing)
58 return current_sino, self.relu(updated_reco)
59
60class Pipeline:
61 def __init__(self, lr, volume_shape, angles,
62 num_detectors, detector_spacing,
63 sdd, sid, voxel_spacing,
64 device, epoches=1000):
65
66 self.epoches = epoches
67 self.model = IterativeRecoModel(volume_shape, angles,
68 num_detectors, detector_spacing,
69 sdd, sid, voxel_spacing).to(device)
70
71 self.optimizer = optim.AdamW(list(self.model.parameters()), lr=lr)
72 self.loss = nn.MSELoss()
73
74 def train(self, input, label):
75 loss_values = []
76 for epoch in range(self.epoches):
77 self.optimizer.zero_grad()
78 predictions, current_reco = self.model(input)
79 loss_value = self.loss(predictions, label)
80 loss_value.backward()
81 self.optimizer.step()
82 loss_values.append(loss_value.item())
83
84 if epoch % 10 == 0:
85 print(f"Epoch {epoch}, Loss: {loss_value.item()}")
86
87 return loss_values, self.model
88
89def main():
90 Nx, Ny = 128, 128
91 phantom_cpu = shepp_logan_2d(Nx, Ny)
92
93 num_views = 360
94 angles_np = np.linspace(0, 2 * math.pi, num_views, endpoint=False).astype(np.float32)
95
96 num_detectors = 256
97 detector_spacing = 1.0
98 voxel_spacing = 1.0
99 sdd = 600.0
100 sid = 400.0
101
102 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
103 phantom_torch = torch.tensor(phantom_cpu, device=device, dtype=torch.float32)
104 angles_torch = torch.tensor(angles_np, device=device, dtype=torch.float32)
105
106 # Generate the "real" sinogram
107 real_sinogram = FanProjectorFunction.apply(phantom_torch, angles_torch,
108 num_detectors, detector_spacing,
109 sdd, sid, voxel_spacing)
110
111 pipeline_instance = Pipeline(lr=1e-1,
112 volume_shape=(Ny,Nx),
113 angles=angles_torch,
114 num_detectors=num_detectors,
115 detector_spacing=detector_spacing,
116 sdd=sdd, voxel_spacing=voxel_spacing,
117 sid=sid,
118 device=device, epoches=1000)
119
120 ini_guess = torch.zeros_like(phantom_torch)
121
122 loss_values, trained_model = pipeline_instance.train(ini_guess, real_sinogram)
123
124 reco = trained_model(ini_guess)[1].squeeze().cpu().detach().numpy()
125
126 plt.figure()
127 plt.plot(loss_values)
128 plt.title("Loss Curve")
129 plt.xlabel("Epoch")
130 plt.ylabel("Loss")
131 plt.show()
132
133 plt.figure(figsize=(12, 6))
134 plt.subplot(1, 2, 1)
135 plt.imshow(phantom_cpu, cmap="gray")
136 plt.title("Original Phantom")
137 plt.axis("off")
138
139 plt.subplot(1, 2, 2)
140 plt.imshow(reco, cmap="gray")
141 plt.title("Reconstructed")
142 plt.axis("off")
143 plt.show()
144
145if __name__ == "__main__":
146 main()