Pixel Discrimination for Gamma Camera Using Artificial Neural Network

1. Ⅰ. Introduction
A. Scintillation crystal and SiPM
Pixel Discrimination Using Artificial Neural Network
for Gamma Camera Detector Module
Daeun Kim, Yong Choi, Kyu Bom Kim, Sangwon Lee, and Donghyun Jang
Molecular Imaging Research & Education (MiRe) Laboratory, Department of Electronic Engineering, Sogang University, Seoul, Korea
Ⅱ. Purpose
Ⅲ. Materials and Methods
Ⅳ. Results
Ⅴ. Summary and Conclusion
Ⅵ. References
C. Artificial neural network
B. Gamma camera system configuration
 LYSO crystal (Sinocera, China)
• 3 mm x 3 mm x 20 mm
• 4 x 4 array
 Silicon photomultiplier (SiPM)
• SPMArray4 (SensL, Ireland)
• Pixel chip size : 3.16 mm x 3.16 mm
• Number of microcells : 3640 per pixel
• Photon detection efficiency : 10 ~ 20 %
 Resistive charge multiplexing circuit
• Array of 100 Ω resistors
• 𝑋 =
𝐵+𝐷
𝐴+𝐵+𝐶+𝐷
, 𝑌 =
𝐶+𝐷
𝐴+𝐵+𝐶+𝐷
D. Training process
A. Crystal position map
 Artificial neural network (ANN) was employed for accurate pixel discrimination and for localization of
the radiation interaction position on the sensor readout by resistive charge multiplexing circuit.
 A new approach was proposed to simplify the training procedure and to optimize ANN structure by
acquiring datasets along a line parallel to x-axis and y-axis.
 Energy resolution and uniformity were measured for the performance evaluation.
 Various pixel discrimination algorithms have been employed to identify the radiation interaction
position in gamma camera detector module. The methods, however, suffer from the nonlinearities and
noise properties deteriorating the discrimination accuracy as the size of detector increases especially
at the edges of the detector [1,2].
 Recently artificial neural network (ANN) has been introduced to identify the radiation interaction
position because of it’s robust capacity compared to other algorithms [3-5].
 However, ANN algorithm usually requires long computational time and training procedure to acquire
pixel by pixel reference data.
 Therefore, structural optimization and procedural simplification are required to practically utilize ANN
algorithm.
M19-24
 Artificial neural network topology [5]
• Input node : 2 (x, y) per class
• Hidden node : 4 per class
• Output node : 1 per class
• Activation function : 𝐹 =
1
1+𝑒−𝑎 , (𝑠𝑖𝑔𝑚𝑜𝑖𝑑)
• Hidden node𝑖 𝑡ℎ
= 𝑊1𝑖 𝑋 + 𝑊2𝑖 𝑌 + 𝑏𝑖
• Output node 𝑗 𝑡ℎ
= 𝑖=1
4
𝑤𝑖𝑗 𝐹𝑖 + 𝑏𝑗
 Training and test datasets
• Pair of x and y (for 12 column and row sets)
Fig. 1. 4 x 4 matrix of LYSO and SiPM
A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
A B
C D
+
Fig. 2. Resistive multiplexing
circuit for 12 x 12 pixels
Fig. 3. Block diagram of gamma camera system
FPGA
PeakDetection
USB3.0
PC
Real-time
Control
Artificial
Neural
Network
12 pixels
12pixels
………
A
B
C
D
Electronics Board
Readout
SiPMArray
(12x12)
…
…
…
…
…
Amplifier
+
A/Dconverter
Na-22 or Cs-137
144:4
DPC
Front-end board DAQ board
SiPM sensors
144:4
DPC
ADC
FPG
A
USB 3.0
Real-time
GUI Control
Artificial
Neural
Network
Processing
Acquisition PC
Fig. 4. Front-end and DAQ boards and acquisition PC
Specification
• 144:4 channel DPC circuit
• 12-bit and 100 MHz ADC
(Analog Devices, AD6644)
• FPGA (Xilinx, Spartan-6)
• USB 3.0 communication
O10
O11
O12
… ...
… ...
Σ Σ Σ Σ
F F F F
Σ
F
W22
W13
W23
W13
W23
w11 w21 w31 w41
Σ Σ Σ Σ
F F F F
Σ
F
X Y
W11
W21
W12
W22
W13
W23
W13
W23
w11 w21 w31 w41
O1
x 12
Independent
network unit
for a class
Fig. 5. Artificial neural network topology
Fig. 6. Training datasets for 12 columns
and 12 rows
(a) (b)
B. Energy resolution
 Source : Na-22, Photo-peak : 511 keV
 Energy spectra of the edge pixels obtained by conventional and proposed method.
(b) Average energy resolution: 15.7 %(a) Average energy resolution: 22.8 %
Fig. 10. (a) Energy spectrum at the edge obtained by the conventional method, (b) energy spectrum at the edge
obtained by the proposed method.
Fig. 9. (a) 3D histogram at
the edge of detector, (b) 3D
flood histogram of 12 x 12
pixels, (c) Profile of
selected white line on (b)
(a) (b)
(c)
C. Counts uniformity and energy resolution obtained by ANN
 Evaluation criteria : Root mean square error : 𝑅𝑀𝑆𝐸 = 𝐸( 𝑚𝑒𝑎𝑛 − 𝑝𝑖𝑥𝑒𝑙 𝑐𝑜𝑢𝑛𝑡(𝑖, 𝑗) 2)
(a) (b)
Mean counts (x104) RMSE (x104) Mean energy resolution (%) RMSE (%)
7.1 2.2 18.0 3.8
Fig. 11. (a) Counts uniformity and (b) energy resolution by proposed method
D. Flood map
Fig. 12. (a) Original flood map and (b) flood map after remapping by ANN
(a) (b)
 Conventional algorithm for crystal position map is challenged by non-uniformity and non-linearity at
the edges of detector modules.
 The use of proposed ANN overcomes these challenges and improves the discrimination accuracy and
energy resolution.
 Since the proposed method is scalable, it is readily applicable to large size detector. Furthermore, the
additional input features could be applied for better positioning accuracy.
Fig. 7. Training dataset and trained results for (a) first column, (b) third
column, (c) first row and (d) third row
(a)
(c)
(b)
(d)
12 pixels
12pixels
RowDataset
12 pixels
12pixels
Na-22 or Cs-137
Column Dataset
1. Training datasets were acquired using a collimated source along a line parallel to x-axis and y-
axis respectively to reduce the training time [6].
2. During ANN training process, each column (or row) dataset produced the probability map in
accordance with it’s distribution on the flood map.
3. After training, all probability maps were combined to make crystal position map.
 Conventional method : Watershed algorithm [7]
 Proposed method : Artificial neural network algorithm
 While the conventional method failed to discriminate adjacent pixels at the edges of scintillation
crystal, ANN successfully discriminates pixels at the edges.
Fig. 8. (a) Crystal
position map
processed by
conventional method,
(b) crystal position
map by the proposed
method
[1] Andras Kufcsak, Maximum likelihood based determination of the position of annihilations in PET detectors. TDK report,
Budapest, 2014, pp. 43-46.
[2] K.A. Stronger et al., “Optimal calibration of PET Crystal position maps using Gaussian mixture models,” IEEE Trans. Nucl.
Sci. 51-1, 2004.
[3] Y. Wang, D. Li, and X. Lu et al., “Self-organizing map neural network-based nearest neighbor position estimation scheme for
continuous crystal PET detectors,” IEEE Trans. Nucl. Sci., vol. 61, no. 5, pp. 2446–2455, Oct. 2014
[4] P. Bruyndonckx, S. Léonard, S. Tavernier, C. Lemaître, O. Devroede, Y. Wu, and M. Krieguer, “Neural network-based position
estimators for PET detectors using monolithic LSO blocks,” IEEE Trans. Nucl. Sci., vol. 51, no. 5, pp. 2520-2525, Oct. 2004.
[5] F.Mateo,R.J.Aliaga,et al., “High-precision position estimation in PET using artificial neural networks,” Nucl. Instr. and Meth.
A604, pp. 366-369, 2009.
[6] H. T. van Dam, S. Seifert, and R. Vinke et al., “Improved nearest neighbor methods for gamma photon interaction position
determination in monolithic scintillator PET detectors,” IEEE Trans. Nucl. Sci., vol. 58, no. 5, pp. 2139-2147, Oct. 2011.
[7] Xiaowen Kang et al, “Comparing Crystal Identification Algorithms for PET Block Detectors”. IEEE Nuclear Science
Symposium Conference Record, 2008.