Digital brain phantoms by generative adversarial network (GAN)

1545 Objectives: Quantitative imaging with PET or SPECT is important for the diagnosis and monitoring of neurological diseases, such as Parkinson’s disease (PD). Machine learning (ML) techniques have been demonstrated to significantly improve the image quality for PET/SPECT, but the development requires a large amount of data for training. Since it is often difficult to collect large population of patient data to support ML research, and the ground truth in clinical data is usually unknown, we have developed a generative adversarial network (GAN) to provide realistic digital brain phantoms that can be used to simulate PET or SPECT data with known ground truth for future ML-based studies in PD. Methods: T1-weighted 3D brain MR images from 192 patients (56% male, 44% female) were segmented using FreeSurfer. The results were reviewed slice by slice on the transaxial plane, and 4,352 2D images that included at least a portion of the striatum were selected. The striatum and the rest of the brain (background) were then assigned realistic activity values to mimic the uptake of radiopharmaceuticals (such as 123I-ioflupane) with a specific uptake ratio ranging from 1 to 9, representing various levels of uptake in a healthy or an affected brain. The corresponding attenuation maps were also generated from the MR images. Activity images and attenuation maps were down-sampled to fit into 128 x 128 frames and were used to train the neural network. A GAN was then developed for generating artificial 2D brain phantoms with the same size as the training images, wherein the discriminator contains six 2D convolutional computations and ends with a sigmoid function, and the generator contains six 2D transposed convolutional computations and ends with a hyperbolic tangent function. The input of the generator is 100 random numbers, and each phantom outputted from the generator contains an activity image and a corresponding attenuation map. Both activity image and attenuation map were input to the discriminator for judgement of their authenticity. The training was performed on a Nvidia P6000 GPU, which ran 3,400 iterations over 3.5 hours. During training, noises were added into the training data to improve the convergences and robustness of the neural networks. Results: The generative network produced effective digital brain phantoms that include activity distributions reflecting the radiopharmaceutical uptake in the brain and corresponding attenuation maps with accurate coefficients. Statistical analysis was performed based on 10,000 generated phantoms using the scores given by the discriminator. The scores showed a normal distribution with a peak approaching 0.5, indicating that the discriminator could not separate the generated phantoms from real training data. The average (Aver) score was 0.4983 and the standard deviation (σ ) was 0.1245. When a threshold Aver-2σ was applied, 98.02% of the generated phantoms had scores above threshold. When a threshold Aver-σ was applied, 82.88% of the generated phantoms having their scores higher than the threshold is considered effective. Conclusion: GANs are a feasible tool for augmenting molecular imaging data in support of ML-based medical imaging and research. The generated digital phantoms can be used in Monte Carlo simulation to generate PET or SPECT data with known ground truth to train ML-based PET/SPECT algorithms for PD studies.