Generation of Digital Brain Phantom for Machine Learning Application of Dopamine Transporter Radionuclide Imaging

Abstract

While machine learning (ML) methods may significantly improve image quality for SPECT imaging for the diagnosis and monitoring of Parkinson's disease (PD), they require a large amount of data for training. It is often difficult to collect a large population of patient data to support the ML research, and the ground truth of lesion is also unknown. This paper leverages a generative adversarial network (GAN) to generate digital brain phantoms for training ML-based PD SPECT algorithms. A total of 594 PET 3D brain models from 155 patients (113 male and 42 female) were reviewed and 1597 2D slices containing the full or a portion of the striatum were selected. Corresponding attenuation maps were also generated based on these images. The data were then used to develop a GAN for generating 2D brain phantoms, where each phantom consisted of a radioactivity image and the corresponding attenuation map. Statistical methods including histogram, Fréchet distance, and structural similarity were used to evaluate the generator based on 10,000 generated phantoms. When the generated phantoms and training dataset were both passed to the discriminator, similar normal distributions were obtained, which indicated the discriminator was unable to distinguish the generated phantoms from the training datasets. The generated digital phantoms can be used for 2D SPECT simulation and serve as the ground truth to develop ML-based reconstruction algorithms. The cumulated experience from this work also laid the foundation for building a 3D GAN for the same application.

Keywords: SPECT; deep learning; generative adversarial network (GAN); phantom.

Figure 1

Digital phantoms to be used to train the GAN. Images are slices containing striatum or a portion of the striatum selected from 3D models. (a) PET (activity) images and (b) generated corresponding attenuation maps.

Figure 2

The training scheme of GAN for generating numerical brain phantoms for the PET/SPECT PD study. Each phantom is composed of two images: one representing the radiopharmaceutical distribution in the brain and the other representing a corresponding attenuation map. The fake phantoms denote the generated image.

Figure 3

The network architecture of the generator (up) and discriminator (low). The number of filters from the first transposed convolution layer to the last in the generator is 1024, 512,256, 128, 64, 2, respectively. Filter size is 4 by 4 in all layers. The number of filters from the first convolution layer to the last in the discriminator is 64, 128, 256, 512, 1024, and 1. Filter size was 5 by 5 in all layers except in the last convolution layer, which was 4 by 4.

Figure 4

Generated phantoms by the developed generator. (a) Generated activity maps and (b) generated corresponding attenuation maps. Each sub-image has 128 by 128 pixels.

Figure 5

The distribution of the phantoms. Yellow bars represent the frequency of generated phantoms, and blue bars represent the frequency of the training phantoms.

Figure 6

The SSIM values. The transverse axis represents the 10,000 generated phantoms. The vertical axis represents the SSIM values when comparing training phantoms with each generated phantom.

Figure 7

Study of the case of the smallest SSIM. (a) Shows the striatal region in a training phantom; (b) Shows the striatal region in a generated phantom; (c) Shows the local SSIM map when comparing (a) and (b). Note (a,b) were from different patients and could be in a different slice position.

Figure 8

Flow chart—training a GAN for generating brain phantoms for SPECT research.