Author: Pouya Azarbar, Nima Kasraie, Peyman Sheikhzadeh 👨🔬
Affiliation: UT Southwestern Medical Center, Shahid Beheshti University of Medical science, Imam Khomeini Hospital Complex,Tehran University of Medical Sciences 🌍
Purpose: Positron Emission Tomography (PET) is crucial for diagnosing and monitoring diseases due to its functional imaging capabilities. However, its high cost, significant radiation exposure, and limited accessibility pose challenges in certain regions. This study aims to develop a deep convolutional neural network (DCNN) for generating virtual PET-like images from CT scans. This approach seeks to provide a complementary tool for faster tumor detection and preliminary diagnosis, improving clinical workflows and accessibility while minimizing radiation exposure.
Methods: This study utilized lung CT scan images and their corresponding PET images from 195 patients, divided into training, testing, and validation sets with a ratio of 70%, 20%, and 10%, respectively. A residual-based U-Net architecture was developed, consisting of an encoder with five down-sampling blocks and a decoder with five up-sampling blocks. Each down-sampling block integrates residual connections, strided convolutions, and batch normalization. The up-sampling process employs transposed convolutions, batch normalization, and Leaky ReLU activation functions to reconstruct the PET-equivalent images.
Results: Developed algorithm achieved SSIM 0.9710, PSNR 30.10 on test dataset and SSIM of 0.9876, PSNR 35.45 on train dataset, train and validation metrics (SSIM, PSNR, MSE, MAE) have similar trend of growing through training process, indicating consistent performance improvements. Results suggest that the algorithm effectively generates PET images, with strong similarity to the ground truth PET images.
Conclusion: The proposed DCNN reliably generates virtual lung PET images from CT scans with high accuracy, offering a valuable tool for early tumor detection and preliminary diagnosis. While not replacing PET imaging, this approach enhances accessibility and reduces costs and radiation exposure. Future work will focus on clinical validation and improving generalizability for widespread adoption.