CT-to-PET Synthesis in the Head-Neck and Thoracic Region via Conditional 3D Latent Diffusion Modeling

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Positron emission tomography (PET) offers crucial physiologic data for cancer staging and treatment evaluation.
• PET availability is limited by cost, radiation, and scanner access.
• Synthesizing PET from computed tomography (CT) is challenging due to the one-to-many mapping between anatomy and tracer uptake.

Purpose of the Study:

• To develop and evaluate a novel conditional 3D latent diffusion framework (3D-LDM) for synthesizing PET images from CT images.
• To improve the efficiency and accuracy of PET synthesis, particularly in the head-neck and thoracic regions.
• To assess the performance of 3D-LDM against existing deep learning methods.

Main Methods:

• A conditional 3D latent diffusion framework (3D-LDM) was proposed for CT-to-PET synthesis.
• The pipeline involved segmenting lungs in CT and encoding PET volumes into a latent space using a 3D autoencoder.
• A conditional 3D diffusion U-Net learned to generate PET latents from CT data through a denoising diffusion process, trained on 900 paired PET/CT studies.

Main Results:

• 3D-LDM achieved superior quantitative fidelity (MAE = 303.05 SUV units, PSNR = 32.64, SSIM = 0.86) compared to transformer, CNN, and GAN baselines (p < 0.001).
• At the lesion level, the model demonstrated high precision (0.76) and recall (0.76), with significantly improved lesion-wise Normalized Mean Square Error (NMSE) of 11.37%.
• The framework successfully synthesized high-fidelity PET images, outperforming state-of-the-art methods in both overall image quality and lesion detection accuracy.

Conclusions:

• The 3D-LDM framework enables efficient and high-fidelity synthesis of PET images from CT scans in the head-neck and thoracic regions.
• The model significantly enhances lesion-level accuracy compared to current state-of-the-art methods.
• While not a substitute for diagnostic PET, 3D-LDM shows potential as a valuable clinical decision support tool in oncology.