Revolutionizing CT Image Quality: AI-Powered Correction for Low Performing Pixels

      Computed Tomography (CT) scans are indispensable tools in modern medicine, offering highly detailed internal views of the human body essential for diagnosing diseases and guiding treatments. However, the precision of these scans can be compromised by inherent hardware limitations, particularly issues with the detector pixels. When these tiny sensors, known as Low Performing Pixels (LPPs), malfunction, they can introduce significant distortions into the final image, rendering critical diagnostic information unusable. These imperfections commonly manifest as "ring" or "streak" artifacts, degrading image quality dramatically.

      Traditionally, addressing severe LPP issues often necessitated the costly and time-consuming replacement of entire detector modules. This not only incurs high service expenses but also prolongs patient waiting times. Recent research from GE HealthCare, as presented in their paper “Low performing pixel correction in computed tomography with unrolled network and synthetic data training” (Source: arXiv:2601.20995), proposes a groundbreaking software-based solution. This innovation leverages deep learning with synthetic data and a novel "unrolled dual-domain" approach to correct LPP artifacts efficiently, bypassing the need for expensive hardware interventions and extensive real-world clinical data for training.

Understanding Low Performing Pixels and Their Impact

      In a CT scanner, X-ray beams pass through the body and are detected by an array of sensitive pixels, which collectively generate raw data known as a sinogram. This sinogram then undergoes mathematical reconstruction, typically using a method like Filtered Back Projection (FBP), to create the final cross-sectional image. Each detector pixel is responsible for collecting signals from a specific angle. If a pixel is "low performing" — meaning its response function is significantly impaired, or it's effectively "dead" – it creates gaps in the sinogram data.

      These gaps translate directly into noticeable ring and streak artifacts in the reconstructed CT image. The closer the LPP is to the scanner's center (isocenter), the more pronounced and disruptive these artifacts become. Such image degradation can obscure vital anatomical details, making it difficult for clinicians to accurately diagnose conditions or plan interventions. This directly impacts patient care and operational efficiency in healthcare facilities.

Limitations of Conventional Correction Methods

      For years, various methods have been employed to mitigate LPP artifacts. Early techniques relied on image processing methods, often involving interpolation to "fill in" missing data. While somewhat effective, these methods had inherent limitations in handling complex artifact patterns.

      With the advent of deep learning (DL), more sophisticated software solutions emerged. Many DL models have been developed to correct artifacts either in the image domain (the final picture) or the sinogram domain (the raw data). Image-based models, such as those employing U-Net, CNNs (like NAFNet), or transformer architectures (like AST), aim to restore the corrupted image directly. Sinogram-based models, including Conjugate-CNN or those using Implicit Neural Representation (INR), focus on interpolating the missing data in the raw sinogram.

      While these deep learning methods have shown promise, they share a significant bottleneck: the reliance on large, meticulously curated datasets of real medical images and their corresponding "ground truth" (artifact-free) versions for training. Collecting such datasets is incredibly expensive and time-consuming, and models trained on data from one scanner or body part often struggle to generalize to different settings, requiring even more data collection. INR-based methods can reduce data requirements but often incur high computational costs during testing, making them impractical for real-world clinical applications requiring rapid image processing.

The Innovation: Dual-Domain Unrolled Network with Synthetic Data

      The GE HealthCare research introduces a novel LPP correction method that tackles these limitations head-on. Their approach is built on two core innovations: the use of synthetic data for training and an unrolled dual-domain network.

      The first major contribution is a natural image-based synthetic data generation strategy. Instead of relying on expensive real clinical data, the researchers generated realistic LPP artifacts on synthetic sinograms derived from natural images. This allows for the creation of vast, unbiased training pairs, where both the corrupted data and the perfect "ground truth" are known. This dramatically reduces the cost and effort associated with data collection, while ensuring the model learns to identify and correct LPP-induced errors effectively across various scenarios.

      The second innovation is a dual-domain unrolled model. This advanced deep learning architecture leverages the intrinsic relationships between the sinogram (raw data) and the image (reconstructed data) domains. An "unrolled network" essentially takes an iterative optimization algorithm (like the Iterative Shrinkage Thresholding Algorithm, or ISTA) and "unrolls" its steps into a deep neural network. This allows the network to learn the optimal parameters for each iteration, effectively combining the strengths of traditional physics-based reconstruction with the powerful pattern recognition capabilities of deep learning. By operating in both domains simultaneously, the model can exploit correlations that single-domain methods miss, leading to more accurate and robust corrections.

      The proposed solution formulates the LPP correction as a compressed sensing problem, where the missing data points in the sinogram (due to LPPs) are recovered iteratively. The ISTA-Net, a CNN-based solver, takes the LPP-corrupted sinogram (reconstructed into a CT image by FBP) and the known LPP locations as input. It then iteratively refines the image, effectively correcting the artifacts.

Key Advantages and Business Impact

      The experimental results from the GE HealthCare study are highly promising, demonstrating that their method significantly outperformed existing state-of-the-art approaches, even when those methods were trained on real CT data. This achievement translates into several critical benefits for healthcare providers and patients:

• Cost Reduction: By eliminating the need for extensive real medical data collection, the costs associated with model training are drastically reduced. Furthermore, software-based correction minimizes the need for expensive detector module replacements, directly lowering maintenance and service expenses.
• Enhanced Image Quality and Patient Safety: Proactive and accurate artifact correction ensures high-quality diagnostic images, reducing the risk of misdiagnosis and improving patient outcomes. This aligns with the fundamental goal of clinical imaging: clear, reliable visualization.
• Operational Efficiency: Reducing the frequency of hardware repairs means less scanner downtime, allowing hospitals and clinics to perform more scans and reduce patient waiting lists. This optimizes the utilization of expensive CT equipment.
• Scalability and Adaptability: The synthetic data training and software-based nature of the solution make it highly adaptable to different CT scanner settings and protocols. This flexibility is crucial for widespread implementation across diverse healthcare environments.
• Privacy Compliance: Using synthetic data for training inherently sidesteps many of the privacy concerns associated with handling vast amounts of sensitive patient data, facilitating deployment without compromising patient confidentiality.

      This research marks a significant step towards more resilient and cost-effective CT imaging systems. By transforming challenging hardware limitations into solvable software problems, it offers a path to consistently superior image quality without the traditional overheads.

Advancing Healthcare Through AI Innovation

      The application of AI and advanced computational techniques to solve critical challenges in medical imaging is a testament to the transformative power of technology. Companies like ARSA Technology are also driving innovation in the healthcare sector, leveraging AI and IoT to enhance various aspects of medical and corporate wellness. For example, the Self-Check Health Kiosk demonstrates how automated, self-service health screening can improve employee welfare and reduce operational burdens in diverse settings, much like the CT LPP correction aims to streamline medical workflows. ARSA’s broader AI Video Analytics capabilities also show how computer vision can provide real-time insights for safety and operational efficiency across various industries, including healthcare.

      This kind of advanced AI optimization and synthetic data generation represents a future where medical technology is not only more powerful but also more accessible, adaptable, and economically viable, benefiting healthcare systems and patients globally.

      To learn more about how advanced AI and IoT solutions can transform your operations and explore potential applications in your industry, we invite you to explore ARSA's range of solutions and contact ARSA for a free consultation.

      Source: Yang, H., Lippenszky, L., Timko, E., Ferenczi, L., & Avinash, G. (2026). Low performing pixel correction in computed tomography with unrolled network and synthetic data training. arXiv preprint arXiv:2601.20995.