TL;DR
This paper introduces a novel "wobbling" method to correct for defective pixels in Cadmium Zinc Telluride (CZT) detectors used in Single Photon Emission Computed Tomography (SPECT) imaging. By acquiring two slightly shifted images of the same object, the method effectively reconstructs data from defective pixels, significantly improving image quality, especially for clustered defects, compared to traditional interpolation techniques.
What Problem Does This Paper Address?
Nuclear imaging techniques like SPECT are crucial for monitoring physiological functions, but the images are often noisy due to low radiotracer amounts and scattering. CZT detectors offer advantages over traditional scintillator detectors due to their superior energy and spatial resolution. However, CZT detectors can suffer from "defective pixels" that exhibit abnormal responses, leading to image artifacts like ring artifacts after reconstruction. These artifacts can degrade image quality and lead to misinterpretations, particularly in small animal imaging where high resolution is critical. Existing methods like interpolation struggle with clustered defective pixels, necessitating a more robust correction approach.
Methods
The core of the proposed method is a "wobbling" technique that involves acquiring two images of the same object with a slight spatial shift (e.g., two pixels horizontally). This creates two "twin images" where the defective pixels in one image correspond to functional pixels in the other.
The process involves:
- Uniformity Correction and Wobbling Path Planning: Initial flood images are used to identify defective pixels and plan the wobbling path.
- Acquisition of Wobbling Images: Two images are acquired with a controlled shift, creating overlapping datasets.
- Registration of Images: A rigid-body transformation is applied to align the two acquired images, accounting for potential rotation and translation.
- Defective Pixel Replacement and Image Fusion: Defective pixels in one image are replaced with their corresponding values from the shifted image. Two fusion methods are explored:
- Method A (Simple Averaging): A straightforward average of pixel values in the overlapping region.
- Method B (Distance-Weighted Function): A weighted average where pixels closer to the boundary of the overlapping region contribute more to the final fused image.
Datasets
The evaluation was performed using phantom experiments. Specifically:
- Flood Phantom: Used for initial uniformity correction and identification of defective pixels.
- Line Source Phantom: Four linear sources of Tc-99m were used to simulate imaging scenarios. The authors deliberately introduced various patterns of defective pixels (2 vertical, 2 horizontal, 3 "L"-shaped, 4 square) into these acquired images for testing.
No new benchmark datasets were introduced; rather, the authors created "bad images" by setting counts to zero in specific pixel locations on top of acquired phantom data.
Key Insights and Ideas
- Leveraging Redundancy: The fundamental insight is that by shifting the detector, information from defective pixels in one acquisition can be found in functional pixels in the other. This creates redundancy that can be exploited for correction.
- Clustered Defects are the Real Challenge: While interpolation works reasonably well for isolated bad pixels, its performance degrades significantly when defects cluster together, forming continuous regions. The wobbling method excels in these challenging scenarios.
- Fusion Matters: The choice of fusion method (simple averaging vs. distance-weighted) impacts the final correction quality, with the distance-weighted approach showing slightly better results in some cases.
- Mechanical Simplicity: The method relies on a simple mechanical stage for shifting the detector, making it potentially adaptable to existing systems.
Key Results
The proposed wobbling method demonstrated significant improvements over conventional interpolation, particularly for clustered defective pixels:
- Quantitative Evaluation:
- PSNR (Peak Signal-to-Noise Ratio): For defective pixels in the Region of Interest (ROI), the wobbling method (both A and B) showed substantial improvements. For instance, with 4 square bad pixels, PSNR increased by 64% (Method A) and 71.9% (Method B) compared to conventional interpolation. The PSNR for wobbling methods was significantly higher, especially as the number of bad pixels increased.
- SSIM (Structural Similarity Index): For background defective pixels, SSIM also showed improvements. For 3 bad pixels, SSIM increased by 11.6% (Method A) and 11.8% (Method B) compared to interpolation.
- Qualitative Evaluation: Line profile analysis showed that the wobbling method recovered missing information more effectively, maintaining counts similar to the reference image, whereas interpolation resulted in a noticeable decrease in counts and information loss around the defective pixel regions.
Limitations and Future Work
- Limitations:
- The paper focuses on correcting "fixed defective pixels" and does not address other defects like geometric deformation, noise corruption, or reconstruction artifacts.
- The current implementation uses a simple mechanical stage, and the accuracy of the correction is dependent on the precision of this stage.
- The method is evaluated on phantom data; real in-vivo imaging performance might introduce additional complexities.
- Future Work:
- Investigating more advanced image fusion algorithms for the wobbling images to further enhance correction effectiveness.
- The authors suggest that the wobbling method is conceptually simple, computationally efficient, and easy to use, making it potentially applicable to other pixelated detectors (Si, GaAs, CdTe) with a suitable motorized system.
Summary updated