Author: Pouya Azarbar, Nima Kasraie, Mahsa Shahrbabki Mofrad, Peyman Sheikhzadeh 👨🔬
Affiliation: UT Southwestern Medical Center, Shahid Beheshti University of Medical science, Imam Khomeini Hospital Complex,Tehran University of Medical Sciences, Tehran University of Medical Science 🌍
Purpose: PET imaging become crucial in diagnosing and managing various diseases, but its key limitation is the lack of detailed anatomical information. Integrating CT-scans with PET images enhances clinical diagnosis by providing complementary anatomical data. However, CT-scans introduce additional radiation exposure, and may not always be available. Generating pseudo-CT from PET scans addresses these issues by reducing radiation exposure, shortening imaging time, and improving workflow efficiency. Deep Convolutional Neural Networks (DCNNs), which have shown promise in generating pseudo-CT from other modalities like MRI, are employed in this study to develop an algorithm for generating brain pseudo-CT from paired non-attenuation corrected PET(NAC).
Methods: 85 patients with paired CT-scans and NAC images from two centres included. The dataset divided into training, validation, and test sets in a 70%-20%-10% ratio. The algorithm was based on a U-Net architecture with convolutional block attention modules (CBAM) integrated into the decoder before upsampling layer(CBAM-UNet). The encoder consisted of 5 down-sampling blocks, while the decoder comprised 5 up-sampling blocks, each beginning with a CBAM. Skip connections were applied between corresponding blocks of the encoder and decoder to preserve essential feature maps.
Results: The algorithm achieved an SSIM of 0.9832 and PSNR of 33.47 on the test dataset, and an SSIM of 0.9897 and PSNR of 36.56 on the training dataset. Additionally, the metrics for the training and validation datasets progressed simultaneously with a similar trend, indicating the absence of both overfitting and underfitting during the training process.
Conclusion: Our results demonstrated that the CBAM-UNet, successfully generated brain CT-scans from NAC images, achieving strong similarity with ground truth images (e.g., SSIM = 0.9832). This approach shows promise for reducing patient radiation dose and improving the quality of attenuation corrected PET images. Further studies are required to evaluate its clinical applicability and effectiveness in real-world settings.