Kernekoncepter
Recent advancements in training robust and generalizable deep learning models can help overcome challenges hindering the deployment of DL medical imaging solutions.
Resumé
This content discusses the use of synthetic data generation and contrastive self-supervised training for central sulcus segmentation in medical imaging. It explores the challenges of segmenting the central sulcus due to its high morphological variability, especially in adolescent cohorts. The study aims to develop robust and adaptable segmentation models using novel approaches to address limited data availability and improve performance on diverse populations.
Structure:
- Introduction:
- Discusses bipolar disorder (BD) and schizophrenia (SZ) as severe mental disorders impacting individuals and society.
- Project Proposal:
- Addresses challenges in central sulcus segmentation due to morphological variability influenced by gyrification changes with age.
- State of the Art:
- Explores different approaches for automatic sulci detection, including feature-based elastic registration, curvature properties analysis, spherical CNNs, BrainVISA software package.
- Synthetic Data Generation:
- Utilizes SynthSeg's generative model to create synthetic images for training robust segmentation models.
- SimCLR Framework:
- Implements contrastive self-supervised learning using SimCLR to pre-train U-Net encoder on synthetic data for improved feature representation.
- Multi-task Learning:
- Combines contrastive pre-training with GM tissue segmentation task to pre-train U-Net model comprehensively.
- Training Strategy:
- Details training parameters, early stopping criterion, loss functions used, batch size, learning rate, validation strategy.
- Quantitative Analysis:
- Evaluates segmentations using Dice similarity coefficient (DSC) and Hausdorff distance metrics.
Statistik
0.7% of the population affected by BD; 1% by SZ (Robinson & Bergen, 2021)
Approximately 55% of children at familial high risk will encounter mental illness in early adulthood (Thorup et al., 2018)