3354 - Delineation of Primary Lung Cancer with Atelectasis Assisted by GANs-Based Synthetic PET Images from CT

B. Dong¹, R. Zheng², X. Sun², M. Chen¹, and Q. Li³; ¹Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Sun Yat-sen University, Guangzhou, Guangdong, China, ²United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ³Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Sun Yat-sen University, Guangzhou, China

Purpose/Objective(s): To prove that virtual PET images synthesized by contrast-enhanced chest CT scan may improve the accuracy of target delineation in lung cancer with atelectasis. Materials/

Methods: Paired PET and enhanced CT images of patients diagnosed with stage I–III lung cancer between 2010 and 2022 were retrospectively obtained from the Sun Yat-sen University Cancer Center. The hypermetabolic regions with SUVmax greater than 2.5 were automatically delineated on PET/CT images by threshold method. We used cascaded Coarse and Fine Multi-task models for contours segmentation of abnormal metabolic activity on enhanced CT, and generate PET images for the whole body as well as tumor regions simultaneously, then obtain the synthetic PET (sPET) image according to the image fusion of the above data. Mean absolute error (MAE), peak signal-to-noise (PSNR) and structural similarity index (SSIM) were applied to evaluate the validity of the image generation methods of Pix2Pix model and Multi-task model. The gross tumor volume (GTV) contours of 20 stage III lung cancer patients with atelectasis based on real PET image combined with CT, enhanced CT only, and sPET image combined with CT, respectively, were evaluated for clinical validation through metrics including Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), average surface distance (ASD), and false positive rate (FPR).
Results: A total of 101 paired imaging data was available for further analysis, of which the training and validation sets contained 69 and 32 cases, respectively. The sPET images of the whole-body generated by Coarse Multi-task model exhibited lower MAE (0.248 vs 0.295 HU), higher SSIM (0.869 vs 0.785), and higher PSNR (32.102 vs 29.538 dB), compared to Pix2Pix model (all p-values < 0.005). In the tumor region, the MAE, SSIM and PSNR of the Coarse Multi-task model were 0.952 HU, 0.144 and 15.469 dB, respectively, while the Fine Multi-task model significantly improved all measurements to 0.521 HU, 0.366 and 21.695 dB, respectively (p-value<0.001). The DSC, HD95, ASD and FPR of primary lesion delineation were significantly improved after the addition of sPET for reference, compared with contrast-enhanced lung CT alone, which were 0.937 vs 0.904, 2.393 vs 5.310 mm, 0.527 vs 1.239 mm and 0.058 vs 0.150, respectively (all p-values < 0.005), while there was no significant difference in FNR of 0.068 vs 0.059 (p=0.253).
Conclusion: Our study demonstrated a deep learning methodology, cascaded multi-task generative adversarial and segmentation network, to synthesize PET images from CT data in lung cancer without radioactive tracer. The model significantly improved the accuracy of target delineation of lung cancer with atelectasis.