TLDR: Researchers have developed a deep learning method that uses a Multiview Ensemble Conditional Diffusion Model to generate synthetic CT images (pseudo-CTs) directly from uncorrected PET scans. This innovative technique accurately corrects for photon attenuation in PET imaging without requiring a separate CT scan, thereby eliminating extra radiation exposure, avoiding image misalignment, and reducing costs. The model processes PET images from three different orientations and employs a voting system to enhance accuracy and remove artifacts, demonstrating strong performance in brain PET scans with minimal error compared to traditional CT-based correction.
Accurate quantification in positron emission tomography (PET) imaging is vital for precise diagnoses and effective treatment monitoring. However, a significant challenge in PET is ‘attenuation,’ which refers to the loss of photons as they pass through biological tissues before reaching the detectors. If not properly corrected, this signal degradation can lead to inaccurate measurements, making it difficult to distinguish between benign and malignant conditions and potentially resulting in misdiagnosis.
Traditionally, this correction is achieved by performing a co-computed Computed Tomography (CT) scan. The CT scan provides structural data, which is then used to calculate how much the photons are attenuated across the body. While effective, this conventional method has several drawbacks: it exposes patients to additional ionizing radiation, carries the risk of spatial misalignment between the PET and CT images, and requires expensive equipment infrastructure.
A Novel Deep Learning Approach
Emerging advancements in neural network architectures offer a promising alternative through the synthesis of artificial CT images, known as pseudo-CT (pCT) images. Researchers have investigated the use of Conditional Denoising Diffusion Probabilistic Models (DDPMs) to generate high-quality pCT images directly from non-attenuation-corrected (NAC) PET images. This innovative method aims to correct attenuation without the need for a separate CT scan.
The core of this new approach lies in a Multiview Ensemble Conditional Diffusion Model. This model leverages all three orthogonal views (transverse, sagittal, and coronal) of the NAC PET images. By combining the DDPM approach with an ensemble voting mechanism, the system generates superior pCT images that exhibit fewer artifacts and improved consistency from slice to slice.
How the Model Works
The proposed architecture utilizes a Denoising Diffusion Probabilistic Model with a UNet backbone. It consists of a noise predictor, which removes noise from CT images, and a condition encoder, which transforms the PET image into a corresponding CT image. The model processes 2D image slices from each of the three orientations (transverse, sagittal, coronal) independently.
A crucial component is the ‘voting process.’ After each of the three models generates its pCT slices, these slices are combined to form three 3D PET images of the same patient. These images then undergo a pixel-by-pixel comparison. If all three values for a given pixel fall within a predefined distance threshold, they are averaged. If only two values meet the threshold, the outlying value is discarded, and the remaining two are averaged. In cases where all three distances exceed the threshold, the two most similar values are averaged. This ensemble voting strategy is designed to detect and eliminate artifacts, ensuring higher quality and more reliable pCT images.
Study and Results
The model was trained using a dataset of 159 paired PET/CT brain images acquired with a Siemens Biograph Vision PET/CT scanner. The dataset included scans using both [18F]-Fluorodeoxyglucose (FDG) and [11C]-Acetoacetate (AcAc) tracers. The images underwent preprocessing steps, including interpolation and normalization, to ensure consistent input dimensions for training.
The study demonstrated both qualitative and quantitative improvements in pCT generation. The method achieved a mean absolute error of 32±10.4 HU on the CT images. More importantly, when comparing PET images reconstructed using the attenuation map from the generated pCT versus the true CT, an average error of only (1.48±0.68)% was observed across all regions of interest in the brain. The voting mechanism significantly improved metrics such as PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index Measure), MAE (Mean Absolute Error), and RMSE (Root Mean Square Error).
Visually, the PET images corrected with the pCT attenuation maps showed almost no discernible difference from those corrected with real CT scans. The approach also proved effective in handling challenges like patient movement between scans, as the pCT is generated directly from the PET image, ensuring perfect alignment. Furthermore, it showed potential in reducing metal streak artifacts caused by dental implants, a common issue in conventional CT imaging.
Also Read:
- Advancing Microwave Imaging with Physics-Guided Generative AI
- A New Approach to Multi-Organ Segmentation with Spatial Prior Guidance
Advantages and Future Directions
This CT-less attenuation correction method offers several significant advantages. It eliminates the need for additional ionizing radiation exposure from CT scans, avoids potential spatial misregistration errors between separate PET and CT acquisitions, and reduces the overall cost associated with requiring two imaging modalities. This makes brain imaging safer, more cost-effective, and more accurate.
While promising, the research acknowledges limitations. The dataset did not include specific pathologies like large tumors or metal plates, and the model’s performance in such scenarios is yet to be fully understood. The voting implementation was relatively simplistic, suggesting potential for further refinement. Additionally, models trained on data from a single scanner may not generalize perfectly to images from different PET scanners due to variations in resolution and noise levels.
Despite these limitations, this work represents a substantial step forward in medical imaging, offering a robust, deep learning-based solution for attenuation correction in brain PET scans. For more detailed information, you can refer to the full research paper: CT-Less Attenuation Correction Using Multiview Ensemble Conditional Diffusion Model on High-Resolution Uncorrected PET Images.
