⚡ electrical engineering

Quantitative comparison of CT-free attenuation correction frameworks of PET imaging in LAFOV PET/CT systems

This study presents a unified, CT-independent evaluation of six CT-free attenuation correction frameworks for LAFOV PET/CT systems, demonstrating that a hybrid approach combining deep learning-generated attenuation maps with MLACF enhancement achieves the highest quantitative accuracy while highlighting the significant impact of external low-attenuating objects on PET quantification.

Original authors: Konstantinos D. Vrakidis, Maqsood M. Yaqub, Mohammadreza Teimoorisichani, Charlotte L.C. Smith, Marijke E. den Hollander, Catharina W. Menke-van der Houven van Oordt, Idris Bahce, Josee M. Zijlstra, G

Published 2026-10-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Konstantinos D. Vrakidis, Maqsood M. Yaqub, Mohammadreza Teimoorisichani, Charlotte L.C. Smith, Marijke E. den Hollander, Catharina W. Menke-van der Houven van Oordt, Idris Bahce, Josee M. Zijlstra, Gerben J.C. Zwezerijnen, Ronald Boellaard

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of medical imaging, doctors often use a technique called positron emission tomography, or PET, to see how the body is functioning at a cellular level. To do this, a patient is given a small amount of a radioactive tracer that lights up areas of high activity, such as tumors. However, as these signals travel through the body to be detected, they are weakened or absorbed by tissues like bone, muscle, and fat. To create an accurate picture, the computer needs to know exactly how much the body is blocking these signals. Traditionally, this has been done by taking a quick X-ray scan, known as a computed tomography, or CT, scan, right before the PET scan. The CT scan provides a map of the body's density, allowing the computer to correct the PET image.

While the radiation from a low-dose CT scan is generally considered safe, it is still an extra dose of radiation that adds to the total exposure a patient receives. This becomes a more significant concern with the newest generation of PET scanners, which are so sensitive that they can produce high-quality images using much less radioactive tracer than before. Because the PET part of the exam is now so gentle, the CT part, which was once a minor contributor to the total radiation, can actually become the largest source of exposure. This has led researchers to ask a critical question: is it possible to create accurate PET images without using a CT scan at all? If they can, they could eliminate the extra radiation entirely, making these scans safer for children, pregnant patients, and people who need frequent monitoring.

A team of researchers at Amsterdam UMC, working with colleagues from Siemens Healthineers, set out to answer this question by testing several different methods for creating these "CT-free" images. They focused on a new type of scanner with a very long viewing area, which allows for the ultra-low-dose imaging mentioned above. The researchers wanted to see if they could replace the CT scan's density map with something else. They tested two main approaches. The first was a physics-based method that used the scanner's own internal background radiation to build a map of the body. The second was an artificial intelligence method that used a computer program to guess the body's density based solely on the uncorrected PET image. They also tested hybrid versions that combined these guesses with advanced mathematical algorithms to refine the results.

To ensure a fair test, the team used data from twelve real patients who had undergone standard PET scans. They did not pick and choose the best cases; they used the raw data exactly as it was, including any movement the patients might have made or any metal implants they might have had. They compared the images produced by their six different CT-free methods against the standard images that used the CT scan as a reference. The goal was to see how closely the new methods matched the traditional ones in terms of accuracy, specifically looking at the brightness of the images in healthy organs and in cancer lesions.

The results showed that the artificial intelligence approach was far superior to the physics-based method on its own. The AI-generated images were much closer to the standard CT-based images, with only small differences in brightness. However, the AI method had a specific weakness: it sometimes missed details like metal implants or air pockets inside the body because it had to guess them from the image. When the researchers combined the AI method with the mathematical refinement algorithm, the results improved even further. This hybrid approach successfully corrected many of the small errors, such as missing metal or air pockets, and produced the most accurate images of all the methods tested.

Despite these successes, the study uncovered a significant problem that affects all CT-free methods: the presence of objects outside the body. Things like pillows, blankets, or clothing that patients wear or lie on during the scan do not show up in the PET image because they do not contain the radioactive tracer. In a standard CT scan, these objects are clearly visible and their effect is corrected. In the CT-free methods, the computer simply does not know they are there. The researchers found that ignoring these items led to a consistent underestimation of the brightness in the final images, with errors averaging around six percent. This suggests that while the technology is getting very good at mapping the body itself, it still struggles to account for the environment around the patient.

The study also highlighted the importance of timing. In some scans, there was a longer gap between the CT scan and the PET scan, which allowed the patient to move slightly. This movement caused the standard CT-based images to be slightly misaligned with the PET data, creating errors in the reference images themselves. The CT-free methods, which are generated from the PET data alone, stayed perfectly aligned with the patient's position. In these cases, the CT-free images were actually more consistent with the patient's true anatomy than the standard method, which was distorted by the movement. This suggests that for patients who cannot hold perfectly still, CT-free methods might offer a more reliable picture, even if the technology is not yet perfect.

Ultimately, the research indicates that a hybrid system, which uses artificial intelligence to guess the body's density and then refines that guess with physics-based math, is the most promising path forward for CT-free PET imaging. It offers the highest accuracy and the best chance of eliminating the extra radiation dose from CT scans. However, the researchers emphasize that for this technology to be ready for widespread clinical use, future developments must find a way to accurately account for external objects like clothing and pillows. Until those objects are handled correctly, they will continue to introduce small but measurable errors into the diagnosis. The work demonstrates that while the dream of a completely radiation-free PET scan is within reach, the final steps require solving the practical details of how patients are positioned and what they wear during the exam.

Technical Summary: Quantitative Comparison of CT-Free Attenuation Correction Frameworks in LAFOV PET/CT

Problem Statement
The emergence of long axial field-of-view (LAFOV) PET/CT systems has enabled diagnostic-quality imaging with significantly reduced injected radiotracer activities. While this reduces the radiation burden from the PET component, it increases the relative contribution of the CT component to the total patient radiation dose. Furthermore, CT-based attenuation correction (AC) remains a source of ionizing radiation and is susceptible to quantitative errors caused by motion mismatches, metal artifacts, and photon starvation. Consequently, there is a growing interest in CT-free PET frameworks. However, existing literature lacks a unified, direct comparison of these frameworks. Previous studies often differ in evaluation methodologies, such as the use of scan-specific CT data for patient bed placement or image registration to correct misalignments, which complicates the assessment of true CT-free performance. Additionally, the impact of low-attenuating objects (LAOs) outside the patient body (e.g., immobilization equipment, clothing) on quantification is frequently unaddressed.

Methodology
This study performed a head-to-head, unified, and CT-independent evaluation of six distinct CT-free attenuation correction frameworks against a CT-based reference on a Biograph Vision Quadra LAFOV PET/CT system.

Data: The study utilized a retrospective dataset of 12 oncological patients who underwent both standard clinical protocols and specific transmission scans. The dataset included two protocols: a short-interval protocol (LD-CT followed immediately by an 8-minute static PET) and a long-interval protocol (LD-CT followed by a 70-minute dynamic PET, with a 5-minute LSO-Tx scan). A separate cohort of 114 patients was used exclusively for training the AI components.
Frameworks Investigated: The six CT-free approaches were grouped into two primary categories, each enhanced by joint-reconstruction algorithms:
1. Physics-Based (LSO-Tx): Utilized the intrinsic background radiation of lutetium oxyorthosilicate (LSO) crystals to generate transmission scans ( $\mu_{LSO-Tx}$ ). These were further processed using Maximum Likelihood Activity and Attenuation (MLAA) or Maximum Likelihood Activity and Attenuation Correction Factors (MLACF) algorithms.
2. Data-Driven (AI/CNN): A 3D attention-gated U-Net synthesized attenuation maps ( $\mu_{CNN}$ ) directly from non-attenuation-corrected (NAC) PET images. These were similarly enhanced with MLAA and MLACF algorithms.
3. Reference: A standard CT-based framework using low-dose CT (LD-CT) reconstructed via Maximum Likelihood Transmission Reconstruction (MLTR).
Evaluation Metrics: Performance was assessed using Standardized Uptake Value mean ( $SUV_{mean}$ ) extracted from volumes of interest (VOIs) in healthy tissues and lesions. Metrics included bias (mean/median difference), variance (standard deviation), Root Sum of Squares (RSS) to combine bias and dispersion, Coefficient of Variation (CV), and Contrast-to-Noise Ratio (CNR).
Specific Investigations: The study explicitly investigated the impact of LAOs by manually inserting CT-based attenuation values for external objects into the CT-free $\mu$ -maps to quantify the bias introduced by their absence.

Key Contributions

Unified Comparison: The study provides a rare direct comparison of physics-based (LSO-Tx), data-driven (CNN), and hybrid (AI + MLAA/MLACF) frameworks using the exact same patient dataset and a strictly CT-independent workflow for the CT-free methods.
Quantification of LAO Impact: The research explicitly quantifies the significant quantification bias introduced by low-attenuating objects (e.g., pillows, clothing) in CT-free frameworks, a factor often overlooked in prior literature.
Hybrid Framework Validation: It demonstrates that combining deep learning-based $\mu$ -map synthesis with physics-based MLACF enhancement yields superior quantitative accuracy compared to standalone approaches.
Motion Sensitivity Analysis: The study highlights how long time-intervals between CT and PET acquisitions (common in dynamic protocols) exacerbate misalignment issues in CT-based references, whereas CT-free methods maintain spatial alignment with the PET emission data.

Results

Quantitative Accuracy: CNN-based frameworks demonstrated substantially lower $SUV_{mean}$ $S U V_{m e an}$ biases compared to LSO-Tx-based approaches.
- Standalone CNN: -9±7% bias (mean ± SD).
- CNN + MLAA: -12±13% bias (mean ± SD).
- CNN + MLACF: -1.6±10.2% bias (mean ± SD).
- LSO-Tx Standalone: 29±29% bias (mean ± SD).
- LSO-Tx + MLAA: 26±100% bias (mean ± SD).
- LSO-Tx + MLACF: 15±15% bias (mean ± SD).
Hybrid Performance: The hybrid framework (CNN + MLACF) achieved the highest agreement with the CT-based reference across all VOIs, with a median difference of -2.6% (-7.3%, 3.4%).
Artifact Correction: While standalone CNN maps often missed internal air pockets and metal implants, the addition of MLACF partially corrected these localized inaccuracies. Conversely, LSO-Tx methods showed overestimation of $\mu$ -values and cross-talk artifacts in high-uptake regions (e.g., bladder, brain), which were partially mitigated by MLACF but not eliminated.
Low-Attenuating Objects (LAOs): The absence of LAOs in CT-free $\mu$ -maps led to a median $SUV_{mean}$ underestimation of approximately 5.9% (3.5%, 8.3%). When CT-based values for these objects were manually inserted, biases were significantly reduced (e.g., to -3.8% for lesions in the CNN framework).
Motion and Alignment: In the long-interval cohort, 70% of scans showed visible PET-CT misalignment. The CT-free frameworks maintained alignment with the PET data, whereas the CT-based reference suffered from misregistration errors, suggesting that CT-free methods may be more robust in dynamic or long-interval scenarios despite the reference bias.

Significance and Claims
The authors conclude that among the investigated frameworks, the hybrid approach combining NAC-PET-based deep learning with MLACF enhancement achieves the highest quantitative accuracy for LAFOV PET imaging. The study emphasizes that while physics-based enhancements can correct gross $\mu$ -map errors (like missing metal or air), the most critical limitation of current CT-free frameworks is the failure to account for low-attenuating objects outside the patient body. The paper asserts that future development and evaluation of CT-free PET frameworks must explicitly address and handle these external objects to avoid substantial quantification biases. Furthermore, the study suggests that emission-only approaches offer inherent advantages in spatial alignment, potentially outperforming CT-based methods in scenarios involving significant patient motion or long acquisition intervals, provided the external object issue is resolved.

Technical Summary: Quantitative Comparison of CT-Free Attenuation Correction Frameworks in LAFOV PET/CT

More like this