Position-Sensitive Silicon Photomultiplier Array with Enhanced Position Reconstruction by means of a Deep Neural Network

This paper demonstrates that applying Deep Neural Networks to a 2x2 array of linearly-graded Silicon Photomultipliers significantly improves position resolution and linearity while increasing the number of resolvable pixels by a factor of 5.7 to 12.1 compared to traditional reconstruction methods.

The Gist

Imagine you are trying to take a photograph of a tiny, glowing firefly landing on a large, dark window. You want to know exactly where on the window it landed.

In the world of medical imaging (like PET scans or SPECT scans), scientists use special sensors called Silicon Photomultipliers (SiPMs) to act as that window. They detect tiny flashes of light created when radiation hits a crystal. The goal is to map exactly where those flashes happen to build a clear picture of what's inside a patient's body.

The Problem: The "Fuzzy" Window

The researchers in this paper were working with a specific type of sensor called an LG-SiPM. Think of this sensor not as a grid of individual pixels (like a standard camera), but as a single, large, continuous sheet of glass with a special "resistive" coating.

When a flash of light hits this sheet, it spreads out, like a drop of ink hitting a wet paper towel. The sensor has only 6 wires (channels) sticking out of it to measure how much "ink" (charge) reached each corner.

The Old Way (Linear Reconstruction):
Traditionally, scientists used a simple math formula to guess where the firefly landed. It's like saying, "If the left wire got 30% of the signal and the right wire got 70%, the firefly must be 30% from the left."

• The Flaw: Real life isn't perfect. The wires have tiny defects, the glass isn't perfectly uniform, and the ink spreads unevenly. This simple math formula gets confused, leading to a blurry, distorted map. It's like trying to navigate a city using a map that has the streets slightly warped; you end up in the wrong neighborhood.

The Solution: The "Super-Intelligent" Translator

The researchers decided to stop using the simple math formula and instead teach a Deep Neural Network (DNN) to solve the puzzle.

Think of the Neural Network as a super-smart translator or a tutor.

1. The Training: They took a laser pointer (acting like the firefly) and moved it to thousands of known, precise spots on the sensor. They recorded what the 6 wires "saw" at every single spot.
2. The Learning: They fed this data into the AI. The AI looked at the messy, distorted signals from the 6 wires and compared them to the true location of the laser. It learned the "quirks" and "glitches" of the sensor. It realized, "Oh, when the top wire reads X and the bottom reads Y, the simple math says it's here, but because of that weird defect in the glass, it's actually there."
3. The Result: Once trained, the AI could look at a new, messy signal and instantly correct the distortion, pinpointing the location with incredible accuracy.

The Magic Numbers: From a Grid to a Forest

The results were shocking.

• The Old Method: Could only distinguish about 540 separate spots on the sensor. Imagine trying to walk through a forest where the trees are very far apart; you can't see the details between them.
• The New AI Method: Could distinguish over 6,500 separate spots!
  • The Analogy: It's like taking that same forest and suddenly realizing every single blade of grass is a distinct tree. The resolution increased by a factor of 12.

Why Does This Matter?

In medical imaging, higher resolution means doctors can see smaller tumors or finer details in the body.

• Before: The sensor was like a low-resolution TV screen (big, blocky pixels).
• After: With the AI, it's like a 4K Ultra HD screen.

The researchers showed that you don't need to build expensive, complex sensors with thousands of wires to get high-definition images. You can use a simpler, cheaper sensor with fewer wires and just let a smart computer do the heavy lifting to fix the blurry parts.

The Bottom Line

This paper proves that by combining a cleverly designed sensor with a "brain" (the Deep Neural Network), we can turn a slightly imperfect, low-channel device into a high-precision imaging tool. It's a perfect example of how modern AI can fix the physical limitations of hardware, making medical scans sharper, clearer, and potentially more affordable.

Technical Summary

Here is a detailed technical summary of the paper "Position-Sensitive Silicon Photomultiplier Array with Enhanced Position Reconstruction by means of a Deep Neural Network."

1. Problem Statement

Gamma-ray imaging applications (such as SPECT and PET) rely on position-sensitive detectors to reconstruct the location of scintillation events. While Silicon Photomultipliers (SiPMs) offer high gain and timing resolution, traditional position reconstruction methods face significant limitations:

• Channel Count vs. Resolution: Standard arrays require one readout channel per pixel, leading to complex electronics. Position-sensitive SiPMs (specifically Linearly-Graded SiPMs or LG-SiPMs) use a resistive network to infer position from fewer channels (e.g., 4 or 6 channels for a 2x2 array), but this introduces non-linearities.
• Distortions and Non-Linearity: The standard "Center of Gravity" (CoG) reconstruction formula assumes ideal, linear sensor behavior. In reality, electronic defects, non-uniformities in the resistive network, and edge effects cause systematic shifts, distortions, and reduced linearity in the reconstructed event maps.
• Limited Granularity: These distortions reduce the number of distinguishable regions ("pixels") within the sensor area, limiting the effective spatial resolution and imaging precision.

2. Methodology

Experimental Setup

• Device: A 2x2 tile of Linearly-Graded SiPMs (LG-SiPMs), each element approximately $8 \times 8 \text{ mm}^2$, creating a total sensitive area of$\approx 16 \times 16 \text{ mm}^2$. The chips use FBK RGB-HD technology with a$20 \mu\text{m}$ cell pitch.
• Readout: A "smart channel" configuration connects the central channels of the 2x2 array, resulting in only 6 readout channels ($Q_1$ to $Q_6$) to infer the hit position.
• Data Acquisition:
  • A blue LED ($470 \text{ nm}$) coupled to an optical fiber was used as a light source.
  • The sensor was moved in a grid pattern ($37 \times 37$steps,$0.5 \text{ mm}$ spacing) using linear stages.
  • At each step, 10,000 waveforms were acquired per channel.
  • Data was filtered by total charge to exclude edge effects and gaps between tiles.

Reconstruction Approaches

The study compared two reconstruction methods:

1. Linear Reconstruction (Baseline):

   • Uses the standard analytical formula derived from the resistive network architecture (Equations 1 & 2 in the paper).
   • Applies a linear transformation (scaling, rotation, and shifting) to map relative coordinates to absolute motor coordinates.
   • Effectively a "zero hidden layer" neural network with fixed or minimally optimized linear weights.

2. Deep Neural Network (DNN) Reconstruction:

   • Architecture: A feedforward network with a 6-unit input layer (normalized charge amplitudes $Q_i$), followed by $N_{layers}$ hidden dense layers (64 units each) using hyperbolic tangent (tanh) activation functions. The output layer consists of 2 units (linear activation) for $x$ and $y$ coordinates.
   • Training Strategy: The dataset was split using three distinct techniques to test generalization:
     • (A) Random split.
     • (B) Chessboard split (alternating grid positions).
     • (C) Random position split.
   • Optimization: Trained using the Adam optimizer to minimize the Mean Squared Error (MSE) between reconstructed positions and the known motor positions. Input charges were normalized by total charge to remove dependence on absolute light yield.

3. Key Contributions

• DNN for Position Correction: Demonstrated that a Deep Neural Network can effectively learn and correct the non-linear distortions inherent in LG-SiPM resistive networks, outperforming the nominal analytical reconstruction formula.
• Granularity Enhancement: Showed that DNNs can significantly increase the effective granularity of the sensor, allowing for the resolution of much smaller areas than previously possible with linear methods.
• Robustness Testing: Validated the method across three different data splitting strategies, proving that the DNN learns the underlying sensor physics rather than just memorizing specific training coordinates.

4. Key Results

Metric	Linear Reconstruction	DNN Reconstruction (Best Case)	Improvement Factor
Mean Resolution ($\sigma$)	$\approx 78.5 \mu\text{m}$	$\approx 67.4 \mu\text{m}$	$\approx 1.17\times$
Mean Systematic Shift ($\nu$)	$\approx 317 \mu\text{m}$	$\approx 41 \mu\text{m}$	$7.8\times$
Granularity ($2\langle d \rangle$)** |$\approx 687 \mu\text{m}$|$\approx 198 \mu\text{m}$| **$3.5\times$
Resolvable Regions ($N$)	$\approx 541$	$\approx 6,530$	$12.1\times$

• Systematic Shift: The most significant improvement was in reducing systematic shifts. The DNN reduced the average displacement error by a factor of 3.4 to 7.8 compared to the linear method, effectively correcting the distortions at tile gaps and borders.
• Resolvable Regions: The number of distinguishable "pixels" increased from ~541 (linear) to ~6,530 (DNN) for a $16 \times 16 \text{ mm}^2$ area, representing a 5.7 to 12.1-fold increase depending on the data splitting method.
• Optimal Architecture: Networks with 3 hidden layers provided the best balance between convergence speed and performance. Adding more layers yielded diminishing returns, while fewer layers (0-2) were suboptimal.
• Resolution: While the intrinsic noise (resolution $\sigma$) improved slightly ($78.5 \to 67.4 \mu\text{m}$), the primary gain was in linearity and the reduction of systematic bias.

5. Significance and Conclusion

• Sub-Millimeter Precision: The study proves that LG-SiPMs, when paired with DNN-based reconstruction, can achieve sub-millimeter spatial resolution ($\approx 200 \mu\text{m}$ granularity) without increasing the number of readout channels.
• Medical Imaging Impact: This technology is highly relevant for compact gamma cameras (SPECT/PET), where reducing channel count lowers cost and complexity while maintaining or improving image quality.
• Future Outlook: While the current study used LED scans (light-only), the authors note that the comparative advantage of DNN over linear methods should hold for real scintillation events. Future work will involve coupling these sensors with scintillators to validate performance under realistic radiation detection conditions, despite the challenges of obtaining ground-truth positions for gamma rays.

In summary, the paper establishes that Deep Neural Networks are a superior alternative to analytical formulas for reconstructing positions in linearly-graded SiPM arrays, effectively transforming a sensor with significant non-linear distortions into a high-precision, high-granularity imaging device.