3528 - High-Fidelity 3D PET Synthesis from CT Using a Mamba-Based Multi-Scale Network: Multi-Center Validation in Lung Cancer
Presenter(s)
X. Li1, J. Qiu2, and Y. Pan3; 1School of Radiology,Shandong First Medical University & Shandong Academy of Medical Sciences, Jinan, China, 2The First Affiliated Hospital of Shandong First Medical University, Jinan, China, 3Department of Radiology, Shandong First Medical University & Shandong Academy of Medical Sciences, Taian, China
Purpose/Objective(s): To develop a clinically adaptable 3D deep learning framework for synthesizing quantitative PET images from routine CT volumes, enabling reliable metabolic assessment for lung cancer while reducing reliance on additional PET scans.
Materials/Methods: We developed a multi-scale 3D UNet integrating selective state-space modeling (Mamba-MS-UNet) to synthesize PET from CT volumes. The Mamba module, embedded in the encoder, efficiently captures long-range 3D spatial dependencies with low computational overhead. A hierarchical SUV-aware loss function was employed to preserve voxel-level quantitative accuracy in tumor regions. The model was trained and internally validated on PET/CT data from Center 1 using standardized preprocessing. External validation was performed on independent datasets from Center 2 and Center 3. At Center 2, volumetric structural fidelity and voxel-wise agreement between synthetic and real PET were evaluated using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). At Center 3, radiomics-based classification of lung adenocarcinoma versus squamous cell carcinoma was performed using (i) real PET + CT + pathology and (ii) synthetic PET + CT + pathology to assess preservation of discriminative metabolic information.
Results:
On Center 2, synthetic PET demonstrated high agreement with real PET, achieving an SSIM of 0.92 ± 0.02 and a PSNR of 32.51 ± 2.49 dB, reflecting high volumetric structural fidelity and strong voxel-level similarity. On Center 3, pathological subtype classification using synthetic PET achieved an AUC of 0.721, which was identical to that obtained using real PET (AUC = 0.721), indicating preservation of clinically relevant metabolic features.Conclusion:
Mamba-MS-UNet enables high-fidelity 3D PET synthesis from CT with improved structural consistency and quantitative SUV reliability. This framework offers a practical low-cost, low-radiation alternative for metabolic imaging in lung cancer and demonstrates strong potential for supporting downstream quantitative analyses and radiotherapy planning.