Deep Learning Models For Stroke Lesion Segmentation And Brain Aging Clocks Liang Shang by Liang Shang instant download after payment.

This thesis presents advanced deep-learning methodologies featuring efficient model architectures and label augmentations to enhance our understanding in both approaches. For voxel-based analysis, we focus on small-stroke lesion segmentation. We introduce the Multi-Stage CrossScale Attention (MSCSA) module, integrated with the U-Net, to enhance small lesion detection by modeling relationships between image patches across different stages and resolutions. Additionally, we propose two novel label augmentations, Multi-Size Labeling (MSL) and Distance-Based Labeling (DBL), which emphasize small lesions and lesion boundaries to improve segmentation accuracy. Both methods demonstrate their effectiveness in detecting the smallest and most subtle structural changes in the human brain.

For brain-wise analysis, we focus on learning brain aging clocks from graph-like resting-state functional connectivity (RSFC) matrices derived from resting-state fMRI. We propose a Two-Stage Graphical Refinement Network (2-SGRN) for brain connectivity age estimation. In the first stage, a quantized prediction is generated using Support Vector Regression (SVR). In the second stage, this prediction is refined using an ensemble of several Graph Neural Networks (GNNs). This approach demonstrates the effectiveness of GNNs and the two-stage framework in capturing essential feature representations, leading to significant insights into brain characteristics.