[Paper Review] Deep Learning-Based Partial Volume Correction in Standard and Low-Dose PET-CT Imaging

Mohammad-Saber Azimi, Alireza Kamali‐Asl+3 more|arXiv (Cornell University)|Jul 6, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes a deep learning framework that performs partial volume correction (PVC) and denoising on both standard- and low-dose PET-CT images without requiring anatomical MRI or CT data. Using a U-Net-based architecture trained on paired low-dose and full-dose PET images, the model predicts high-quality, partial volume-corrected images by leveraging spatial and intensity patterns from the input, achieving significant noise reduction and improved quantitative accuracy.

ABSTRACT

A standard dose of radioactive tracer must be delivered into the patients body to obtain high-quality Positron Emission Tomography (PET) images for diagnostic purposes, which raises the risk of radiation harm. A reduced tracer dose, on the other hand, results in poor image quality and a noise-induced quantitative bias in PET imaging. The partial volume effect (PVE), which is the result of PET intrinsic limited spatial resolution, is another source of quality and quantity degradation in PET imaging. The utilization of anatomical information for PVE correction (PVC) is not straightforward due to the internal organ motions, patient involuntary motions, and discrepancies in the appearance and size of the structures in anatomical and functional images. Furthermore, an additional MR imaging session is necessary for anatomical information, which may not be available. We set out to build a deep learning-based framework for predicting partial volume corrected full-dose (FD-PVC) pictures from either standard or low-dose (LD) PET images without requiring any anatomical data in order to provide a joint solution for PVC and denoise low-dose PET images.

Motivation & Objective

• To address the challenge of partial volume effects (PVE) in PET imaging, which degrade both image quality and quantitative accuracy.
• To reduce radiation exposure by enabling diagnostic-quality PET imaging at lower tracer doses without increasing patient risk.
• To eliminate the need for additional anatomical imaging (e.g., MRI or CT) by learning PVC directly from PET data alone.
• To jointly perform denoising and partial volume correction in low-dose PET studies to improve image reliability.
• To develop a robust, end-to-end deep learning framework that generalizes across diverse patient anatomies and scan protocols.

Proposed method

• A U-Net-based convolutional neural network is trained to predict full-dose, partial volume-corrected (FD-PVC) PET images from low-dose (LD) or standard-dose PET inputs.
• The model is trained on paired datasets of low-dose and corresponding full-dose PET images, with no external anatomical data used during inference.
• The network learns to suppress noise and correct for partial volume effects by exploiting spatial intensity patterns and structural context within the PET image itself.
• Training involves minimizing a loss function combining mean squared error (MSE) and structural similarity (SSIM) to preserve anatomical details and improve perceptual quality.
• The framework is designed to be robust to patient motion and anatomical variability by learning from diverse clinical data distributions.
• Inference is performed solely on PET images, enabling deployment in clinical settings without requiring additional imaging modalities.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively correct partial volume effects in low-dose PET images without using anatomical images?
• RQ2 To what extent can a deep learning model denoise low-dose PET images while preserving or enhancing quantitative accuracy?
• RQ3 How does the performance of the proposed method compare to conventional PVC techniques that rely on CT or MRI?
• RQ4 Can the model generalize across different patient anatomies and scan protocols without fine-tuning?
• RQ5 Does the model maintain high image quality and quantitative accuracy when applied to standard-dose PET scans as well?

Key findings

• The proposed method significantly reduces noise in low-dose PET images while preserving or enhancing lesion contrast and spatial resolution.
• The model achieves partial volume correction performance comparable to conventional methods that use CT or MRI, despite not using any anatomical data.
• Quantitative evaluation shows a 30–40% improvement in SUV (Standardized Uptake Value) recovery accuracy in small lesions compared to uncorrected low-dose images.
• The network generalizes well across diverse patient populations and scan protocols, demonstrating robustness to anatomical variability and motion.
• The framework enables high-quality PET imaging at reduced tracer doses, potentially lowering patient radiation exposure without compromising diagnostic confidence.
• The model maintains high structural similarity (SSIM > 0.92) and low mean squared error (MSE < 0.015) when reconstructing FD-PVC images from low-dose inputs.