Methods and systems for generating PET images from CT images using a biodiffusion model.

Abstract

The invention proposes a method and system for synthesizing positron emission tomography (PET) images from computed tomography (CT) images using a deep learning model. The solution includes a process for constructing a large-scale training dataset from deidentified CT–PET image pairs and automatically spatially paired, combined with the generation of guidance data including attenuation maps and attention maps. The deep learning model is trained based on the Brownian sphere diffusion mechanism in latent space, where image generation is conditioned by the latent representation of the CT image along with the aforementioned medical knowledge information. Simultaneously, a registration adjustment mechanism is integrated into the training process to predict and correct positional discrepancies between the generated image and the actual image through the deformation field, thereby improving the accuracy of anatomical structure.

Images