Spatial-temporal cascaded network for dynamic [ 11 C]acetate cardiac PET parametric images generation based on one-tissue compartment model

• Published in: Medical physics (Lancaster) Vol. 52; no. 8; p. e18016
• Main Authors: Liu, Shuai, Gong, Tan, Shi, Ximin, Huo, Li, Shang, Fei
• Format: Journal Article
• Language: English
• Published: United States 01.08.2025
• Subjects: Acetates; Carbon Radioisotopes; Deep Learning; Heart - diagnostic imaging; Humans; Image Processing, Computer-Assisted - methods; Kinetics; Positron Emission Tomography Computed Tomography; Time Factors; one‐tissue compartment model; cardiac PET; deep learning
• ISSN: 2473-4209
• DOI: 10.1002/mp.18016

One-tissue compartment model (1TCM) kinetic parameters calculated from dynamic [ C]aceta te cardiac PET/CT imaging can assess cardiac function and assist clinical diagnosis. However, the long acquisition time of dynamic data hinders its clinical application. This study proposed a deep learning-based method for the generation of [ C]acetate 1TCM kinetic parametric images with shortened dynamic PET data, aiming to explore the feasibility of reducing the time required for parametric analysis. A spatial-temporal cascaded network (STCN), consisting of two convolutional modules and one Transformer module, was proposed to generate parametric images K , k , and v . The STCN was trained and tested on [ C]acetate dataset (training/testing: 40 subjects/17 subjects) using 10 frames of dynamic data acquired within the first 10 min of scanning. The parametric images fitted from 40 min of dynamic data using non-linear least squares (NLLS) are considered the reference standard (RS). A temporal loss was incorporated into the training process by integrating the kinetic model. The performance of the STCN was evaluated using normalized root-mean-square error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM). Weighted Akaike information criterion (WAIC) and coefficient of variance (CoV) were calculated within the myocardial region to evaluate the model's goodness-of-fit and the parameter's degree of variability. The proposed method was compared with NLLS and multilinear least squares fitted on 10 min of dynamic data (CM_10 and MLM_10). Three deep learning-based methods, that is, U-Net, Pix2pix, and CycleGAN, were also trained for comparison. Furthermore, ablation experiments were performed to assess the contribution of individual components of the STCN to the generation of parametric images. The STCN achieved the best PSNR and SSIM for k and v parametric images (PSNR: 25.718 ± 2.635 and 32.230 ± 4.090; SSIM: 0.864 ± 0.056 and 0.944 ± 0.041, respectively). The PSNR for the K images generated by STCN was lower than that generated by the Pix2pix model (28.927 ± 2.956 vs. 28.930 ± 2.705). The 1TCM parameters obtained by STCN achieved an average WAIC of 635.64 ± 38.44 in the myocardial region. No significant difference in CoV within the myocardium was found between RS and parametric images derived from STCN. The ablation study results demonstrated that our proposed model architecture and specialized loss functions could improve the quality of the generated parametric images in NRMSE, PSNR and SSIM. The result of the present study shows that the proposed STCN can generate 1TCM parametric images using only 10 min of dynamic [¹¹C]acetate PET data, demonstrating its potential for calculating cardiac [ C]acetate PET 1TCM kinetic parameters in clinical practice.