Author: Rashmi Bhaskara, Shravan Bhavsar, Ananth Grama, Oluwaseyi Oderinde, Shourya Verma 👨🔬
Affiliation: Purdue University 🌍
Generating Synthetic Positron Emission Tomography from Computed Tomography using Lightweight Diffusion Model for Head and Neck Cancer
Purpose: To generate synthetic PET tumor avidity segments directly from CT images, with the aim of providing insights into tumor biochemical activity to support clinical decision-making, particularly in resource-limited settings worldwide.
Methods: This study developed a diffusion-based generative model to map CT segments associated with tumor masks to corresponding PET segments. A paired CT-PET dataset, with tumors delineated based on a 40% threshold of the maximum standardized uptake value (SUVmax), was utilized for training in patients with head and neck cancer. The tumor mask on CT served as a prerequisite for synthesizing tumor PET avidity segments. The synthetic PET outputs were assessed for pixel-wise similarity to ground-truth PET data, quantified using peak signal-to-noise ratio (PSNR), mean squared error (MSE), and mean absolute error (MAE).
Results: The generated synthetic PET segments showed promising fidelity to reference PET within the GTV contours. A reasonable synthesis of the original PET segment was indicated by the quantitatively acceptable PSNR, MSE, and MAE values of 23.13 dB, 0.037431 and 0.106769, respectively. To fully match the visual quality and observer performance of real PET scans, more enhancements are required, according to qualitative inspection. This suggests that future work will focus on improving the model's training and hyperparameter tuning.
Conclusion: This study shows the feasibility of generating synthetic tumor PET avidity directly from CT images via a lightweight diffusion-based model for patients with HN cancer. Further study will focus on enhancing the performance of the model.