Anisotropic to Isotropic Reconstruction of Infant Brain MRI using Deep Learning

Sammanfattning

Brain magnetic resonance imaging (MRI) is an important tool for diagnosing neurological diseases and is widely used in research to study brain development and aging. However, acquiring high-resolution (HR) isotropic MRI is time- and resource-consuming, which often leads to acquisition of low-resolution (LR) anisotropic MRI instead. In recent years, deep learning has shown promising advancements in synthesizing isotropic MRI from anisotropic scans. However, the models are often trained for certain tasks, and may need re-training or re-optimization to perform well in other scenarios. Consequently, this thesis aims to construct a deep learning pipeline specifically optimized for anisotropic to isotropic reconstruction of T2-weighted (T2w) infant brain MRI. The proposed approach is a supervised, 3D patch-based residual U-Net architecture, which has been carefully optimized in terms of network depth, patch size, input channels, residual units, and data augmentation. The final model outperforms a baseline interpolation-based framework, both with visual assessment and quantitative measures. However, the network presents limited generalization to unseen types of LR anisotropic input. To address this limitation, augmentation is introduced by adding more types of LR anisotropic MRI to the training data. This shows promising results and can, with further development, be utilized for research use, and eventually in clinical practice. In conclusion, this work successfully implements a deep learning pipeline for anisotropic to isotropic reconstruction in T2w infant brain MRI through systematic optimization of model hyperparameters.