Maximum-Likelihood--Based Position Decoding of Laser Processed Converging Pixel CsI: Tl Detectors for High-Resolution SPECT

This study demonstrates that a novel converging-pixel CsI:Tl detector fabricated via laser-induced optical barriers, when combined with maximum-likelihood decoding and validated by precise pencil-beam experiments, achieves high spatial resolution and energy performance suitable for advanced SPECT systems.

The Gist

The Big Picture: Making Medical Eyes Sharper

Imagine a doctor trying to take a picture of a tiny tumor inside a patient's body using a special camera that sees radiation (called a SPECT scanner). The problem is, the camera is a bit blurry. It's like trying to read the fine print on a medicine bottle while wearing thick foggy glasses. The doctor can see the general area, but not the tiny details.

This paper is about building a super-sharp lens for that camera. The researchers created a new type of "digital retina" (a detector) that can pinpoint exactly where radiation hits, allowing for much clearer, higher-resolution images of the human body.

---

1. The Problem: The "Foggy" Detector

Traditional detectors are like a grid of square tiles. When a particle of radiation hits the detector, the light bounces around, and the computer has to guess which tile it hit.

• The Issue: If the tiles are too small, they are hard to make. If they are too big, the image is blurry.
• The Old Way: Making these tiny tiles usually involves cutting the crystal with a saw and gluing reflective tape between them. It's like trying to cut a block of cheese into perfect tiny cubes by hand—it's messy, expensive, and you lose some of the cheese (sensitivity) in the gaps.

2. The Innovation: Laser "Origami"

Instead of cutting the crystal, the researchers used a laser to "draw" invisible walls inside the crystal.

• The Analogy: Imagine a giant block of Jell-O. Instead of cutting it with a knife, you use a laser to create invisible, hard lines inside the Jell-O that guide the light.
• The Result: They created a Converging-Pixel design. Think of this like a funnel or a megaphone.
  • The top of the crystal (where the radiation enters) has small openings (1.6 mm).
  • The bottom of the crystal (where the camera sees it) has wider openings (2.0 mm).
  • Why? This funnels the light toward the camera, making the signal stronger and easier to read, just like a megaphone makes your voice louder.

3. The Challenge: The "Where Did It Land?" Game

Now that they have this special crystal, they need a way to tell the computer exactly which "funnel" the radiation hit.

• The Old Method (CoG): This is like a seesaw. If a heavy weight lands on the left side, the seesaw tips left. The computer guesses the position based on how much the "seesaw" (the light) tips.
  • Flaw: It works okay in the middle, but near the edges, the seesaw gets confused and the guess is wrong.
• The New Method (Maximum Likelihood - ML): This is like a detective with a cheat sheet.
  • Before the experiment, the researchers "trained" the computer. They shot a tiny laser beam at every single spot on the crystal and recorded exactly how the light looked for each spot.
  • When a real radiation hit happens, the computer doesn't just guess; it looks at its "cheat sheet" (the database) and says, "This light pattern matches exactly the spot at coordinates X, Y."
  • It uses math to find the most likely answer, even if the signal is noisy.

4. The Experiment: The "Pixel Hunt"

To test this, they built a robot arm that could move a tiny radioactive beam (a "pencil beam") to 625 different spots on the crystal.

• The Test: They fired the beam at every spot and asked the computer: "Where did I hit?"
• The Results:
  • Old Method (CoG): It could only clearly see the middle 15x15 spots. The edges were a blurry mess.
  • New Method (ML): It could clearly identify all 25x25 spots, even at the very edges and corners.
  • Accuracy: The new method was off by less than 1 millimeter on average. That's like hitting a bullseye on a dartboard from across the room.

5. The "Interpolation" Secret Sauce

The researchers tried three different ways to fill in the gaps in their "cheat sheet" (mathematical interpolation):

1. Bilinear: Like connecting dots with straight lines. (Okay, but a bit blocky).
2. Bicubic: Like drawing smooth curves between dots. (Better, very smooth).
3. Nearest-Neighbor: Like snapping the answer to the closest known dot. (Surprisingly, this was the winner! It was the most precise and easiest to sort).

The Takeaway

This paper proves that by combining laser-fabricated funnels (the crystal) with smart detective math (the ML algorithm), we can build medical cameras that see much smaller details.

Why does this matter?
If a doctor can see smaller lesions (tiny tumors) earlier, they can treat them sooner. This technology could lead to:

• Clearer images of the brain and heart.
• Lower radiation doses for patients (because the detector is so sensitive it needs less radiation to see clearly).
• Faster scans (less time in the machine).

In short, they turned a blurry, foggy medical camera into a high-definition one, using lasers and smart math.

Technical Summary

1. Problem Statement

Single-Photon Emission Computed Tomography (SPECT) is a vital nuclear medicine modality but suffers from limited spatial resolution compared to PET. Current clinical systems typically achieve intrinsic spatial resolution below 5 mm. Traditional scintillator arrays rely on mechanical pixelation (cutting and polishing crystals with reflectors), which becomes labor-intensive and expensive at sub-millimeter scales. Furthermore, the inclusion of reflective material reduces the packing fraction and detection sensitivity. While continuous scintillator crystals offer higher sensitivity, accurately decoding the interaction position of $\gamma$-photons within them remains a challenge. Existing decoding algorithms, such as the Center-of-Gravity (CoG) method, often suffer from limited accuracy and pronounced edge effects, while statistical methods require extensive calibration data. Additionally, there is a need for detector geometries that can simultaneously enhance spatial resolution and sensitivity.

2. Methodology

A. Detector Fabrication (LIOB Technique)
The study utilized a Laser-Induced Optical Barrier (LIOB) technique to fabricate a novel converging-pixel architecture in a CsI:Tl scintillator crystal ($50 \times 50 \times 8.05$ mm³).

• Geometry: Instead of parallel pixels, the laser inscribed optical barriers to create converging pixels. The entrance face (interaction side) has a pixel pitch of 1.6 mm, which expands to 2.0 mm at the photodetector interface. This geometry guides scintillation light toward the photosensors, improving collection efficiency.
• Process: Tightly focused, high-intensity laser pulses created localized refractive index changes (optical barriers) without mechanical segmentation, achieving a 100% fabrication yield for a $25 \times 25$ pixel array.
• Readout: The crystal was coupled to an $8 \times 8$ Multi-Pixel Photon Counter (MPPC) array via a light guide. Signals were processed by a custom 64-channel analog front-end and digitized by a CAEN V1740 waveform digitizer.

B. Experimental Setup

• Calibration: A custom-built four-axis motion platform was developed to deliver a finely collimated pencil beam (1 mm diameter) from a $^{99m}$Tc source (140 keV) across a $25 \times 25$ grid of positions. This generated 625 distinct datasets for training and validation.
• Validation: A flood-irradiation experiment was conducted using a $^{57}$Co source (122 keV) to test performance under broad-beam conditions.

C. Decoding Algorithms
Two primary algorithms were implemented and compared:

1. Center-of-Gravity (CoG): A non-calibration method calculating the weighted average of light distribution. A noise-suppression strategy (subtracting the $N$-th largest signal) was applied.
2. Maximum Likelihood (ML): A statistical approach utilizing a Mean Detector Response Function (MDRF) derived from the collimated calibration data. The ML algorithm maximizes the log-likelihood function to determine the most probable interaction pixel.
   • Interpolation Schemes: To refine the MDRF from the coarse calibration grid to a fine decoding grid, three interpolation methods were tested: Bilinear, Bicubic, and Nearest-Neighbor.

3. Key Contributions

• Novel Architecture: First demonstration of a converging-pixel CsI:Tl detector fabricated entirely via the LIOB laser technique, designed specifically to enhance light collection and spatial resolution for SPECT.
• Algorithmic Optimization: Comprehensive evaluation of ML-based decoding with different interpolation schemes for converging geometries, identifying Nearest-Neighbor interpolation as the superior method for discrete pixel classification.
• Robustness Validation: Demonstrated that the ML-based decoding remains effective across different photon energies (122 keV vs. 140 keV) and under flood-irradiation conditions, proving energy independence.
• Noise Handling: Showed that the ML method inherently handles electronic noise and edge effects better than CoG, rendering additional noise-thresholding strategies less critical for this specific detector configuration.

4. Results

Energy Resolution

• Flood Irradiation: 15.76%.
• Collimated Experiment: 11.79 ± 0.53% (average across 625 positions).

Position Decoding Performance

• CoG Method:
  • Achieved a Peak-to-Valley Ratio (PVR) of 1.91 in the central region.
  • Only resolved approximately $15 \times 15$ pixels clearly; edge and corner pixels were merged and indistinguishable.
• ML Method (Bicubic Interpolation):
  • Resolved all $25 \times 25$ pixels.
  • Achieved a PVR of 6.09 (256 resolution) to 6.47 (512 resolution) in collimated data.
  • Positional Error: Average error of 1.07 ± 0.45 mm across the array.
• ML Method (Nearest-Neighbor Interpolation):
  • Produced discrete, unambiguous decoding results for all $25 \times 25$ positions.
  • Positional Error: Achieved the best overall accuracy with an average error of 1.00 ± 0.42 mm.
  • Minimum error at the center: 0.18 mm; Maximum error at edges: 1.01 mm.
  • Successfully resolved distinct peaks in flood-irradiation profiles where CoG failed.

Comparison
The ML method with Nearest-Neighbor interpolation improved the PVR in the central region by 51.31% compared to CoG (from 1.91 to ~2.89–3.17 in flood data) and provided complete pixel resolution even at the crystal periphery, where CoG failed.

5. Significance

This study validates the feasibility of combining laser-processed converging-pixel scintillators with Maximum-Likelihood decoding to overcome the spatial resolution limits of current SPECT systems.

• Performance: The approach achieves sub-millimeter positioning accuracy (1.00 mm average error) and high energy resolution, enabling the detection of smaller lesions.
• Efficiency: The converging geometry improves photon collection efficiency without sacrificing packing fraction, addressing the sensitivity-resolution trade-off inherent in traditional collimators.
• Scalability: The 100% fabrication yield of the LIOB technique suggests a viable path for mass-producing high-performance, high-resolution detectors for clinical PET and SPECT applications.
• Future Impact: The robustness of the ML algorithm across different radionuclides and its ability to handle edge effects make it a strong candidate for next-generation medical imaging systems, potentially reducing acquisition times and administered radiation doses.