Brain phantom generation by generative adversarial net (GAN) for AI-based emission tomography

点击翻译按钮获取中文摘要

Project Summary Positron emission tomography (PET) and single photon emission computed tomography (SPECT) are useful functional medical imaging techniques that can be performed to evaluate brain functions such as regional cerebral perfusion and neurotransmission. The spatial resolution of reconstruction for PET is usually 3-6 mm, and for SPECT is only 1-2 cm. Motivated by the latest advances in artificial intelligence (AI)/machine learning (ML) and its successful application to MRI and CT, it is highly desirable to develop an ML-based system for PET/SPECT cerebral image reconstruction (one of our specific interests is Parkinson disease) to achieve higher resolution and lower noise than using conventional approaches. However, to develop such a learning system, ground-truth data (accurate images, used as the labels) that guide the training are unavailable from the the real world. Published ML systems for PET imaging have used reconstructed images from conventional methods as the label to guide the training. As a result, the goal was only targeted to improve reconstruction speed, rather than improving the image quality. Since the quality of reconstructed image by ML system cannot exceed the guiding images, the performance of ML system cannot surpass conventional methods. Therefore, in this project, we propose a two-year project that will use conditional generative adversarial networks (GAN) to generate digital 2-D human brain phantoms, which will be highly similar to real human brains. The generated phantoms will serve as the (precise) ground-truth data to develop ML-based PET/SPECT reconstruction systems (our future research). The generated phantoms will contain an activity image and an attenuation map. Hence, results from this work can be used for simulating brain PET or SPECT examinations for various neurological disorders, and neural network can be trained with known ground truth. In addition, designing ML systems often relies on large amounts of data, but it is not easy to access data from a large number of patients in the US for specific medical research (mature ML systems developed for computer vision and image classification often involve images on the million level for training). Existing ML systems developed for MRI, CT, and PET imaging often merely uses a few tens of patient data for training and even less data to validate. Therefore, those systems are high-likely overfitted to the data used in training. With the generation system proposed from this project, we can produce a large phantom population to avoid the overfitting problem when design the AI image-reconstruction system. Once the GAN system is successfully developed, it can be easily transplanted to phantom generation for the AI-based CT and AI-based MRI. The method is also potentially extendable to generate phantom populations of torso, abdomen, and extremities for simulating cardiac imaging, tumor imaging, etc.