PTNet: A High-Resolution Infant MRI Synthesizer Based on Transformer
This work addresses data scarcity in infant neurodevelopment research by providing a more stable and efficient synthesis method, though it is incremental as it builds on existing generative approaches.
The paper tackles the problem of synthesizing high-resolution infant MRI scans to address data scarcity due to collection challenges, introducing PTNet, a transformer-based framework that outperforms CNN-based GAN models in accuracy and model size without adversarial training.
Magnetic resonance imaging (MRI) noninvasively provides critical information about how human brain structures develop across stages of life. Developmental scientists are particularly interested in the first few years of neurodevelopment. Despite the success of MRI collection and analysis for adults, it is a challenge for researchers to collect high-quality multimodal MRIs from developing infants mainly because of their irregular sleep pattern, limited attention, inability to follow instructions to stay still, and a lack of analysis approaches. These challenges often lead to a significant reduction of usable data. To address this issue, researchers have explored various solutions to replace corrupted scans through synthesizing realistic MRIs. Among them, the convolution neural network (CNN) based generative adversarial network has demonstrated promising results and achieves state-of-the-art performance. However, adversarial training is unstable and may need careful tuning of regularization terms to stabilize the training. In this study, we introduced a novel MRI synthesis framework - Pyramid Transformer Net (PTNet). PTNet consists of transformer layers, skip-connections, and multi-scale pyramid representation. Compared with the most widely used CNN-based conditional GAN models (namely pix2pix and pix2pixHD), our model PTNet shows superior performance in terms of synthesis accuracy and model size. Notably, PTNet does not require any type of adversarial training and can be easily trained using the simple mean squared error loss.