Medical Image Translation from PET to CT using Deep Learning Techniques

Abstract

Medical imaging plays a crucial role in the early detection, diagnosis, and treatment planning for various diseases, particularly cancer. Positron Emission Tomography (PET) and Computed Tomography (CT) are two common imaging techniques for cancer detection and tumor characterization. PET gives information about the metabolic activity and functional data, while CT gives detailed anatomical structure. The combination of both modalities improves the accuracy of diagnoses, but it also has some downsides, such as more radiation exposure and higher costs. To address this limitation, we propose an image translation framework for synthesizing CT images directly from PET scans using advanced deep learning techniques. This work examines three deep learning models specifically: a Conditional Generative Adversarial Network (cGAN), a Conditional Diffusion Model, and a Hybrid Diffusion-GAN architecture. The models were trained and evaluated using the QIN-Breast dataset, which contains paired PET and CT images. Quantitative evaluation metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were used to evaluate the performance of the models. The findings indicated that the GAN-based model provided the best reconstruction quality, achieving a PSNR of 32.55 dB and an SSIM of 0.8799. The Diffusion model, while producing smooth and anatomically consistent images, resulted in a PSNR of 15.46 dB and an SSIM of 0.776. The Hybrid model effectively combined the strengths of both approaches, producing a PSNR of 21.48 dB and an SSIM of 0.833. This study shows the capacity of deep learning to improve medical imaging workflows by generating synthetic CT generation from PET data, thus reducing patient radiation exposure and associated healthcare costs. Future research will examine the extension of these models to 3D volumetric reconstructions and their integration into clinical workflows