Abstract:Dual-tracer positron emission tomography (PET) has emerged as a promising nuclear medical imaging technology. It reflects more information about biological functions than only using single tracer, which is of great value for disease research, clinical diagnosis and treatment. Furthermore, image reconstruction and segmentation often are treated as two sequential problems. This paper integrates these two problems into a coherent framework. In recent years, deep learning has attracted growing interests in medical image processing due to its powerful feature extraction ability. The reconstruction of simultaneous dual-tracer activity map based on neural network has shown an advantage of not relying on prior information and interval injection that conventional methods always need. In this paper, we propose a joint deep neural network for reconstruction of two individual tracer and segmentation by clustering of time activity curves. In the part of reconstruction, a classic Generative Adversarial Networks (GAN) called pix2pix is introduced as the basic framework. A 3D U-Net is used as the generator realize reconstruction. Then the time activity curves (TACs) in the reconstruction results are extracted and used for clustering to achieve segmentation. The loss of segmentation is used to guide the training of reconstruction network in return. The performance of segmentation network exceeds that of the previous joint method and the performance of reconstruction network is further improved.

Deep-learning Based Joint Estimation of Dual-Tracer PET Image Activity Maps and Clustering of Time Activity Curves