Skip to content

07-brain-segmentation.md

Latest commit

 

History

History
118 lines (82 loc) · 7.61 KB

07-brain-segmentation.md

File metadata and controls

118 lines (82 loc) · 7.61 KB
Raw

VII. Brain Segmentation & Lesion Detection

← Back to Wiki Index

Whole-Brain Segmentation

Whole-brain segmentation partitions the brain into anatomical regions — cortical parcels, subcortical structures, and white matter tracts — enabling quantitative analysis of brain structure.

Papers

  • Whole Brain Segmentation — Fischl et al., 2002

    • Key idea: Probabilistic atlas-based segmentation using spatial priors and intensity models in a Bayesian framework with MRF regularization.
    • 🔗 Download

  • SynthSeg: Segmentation of Brain MRI Scans on Synthesis Images — Billot et al., 2023

    • Key idea: Trained entirely on synthetic images with random contrasts, achieving segmentation that works on any MRI scanner or acquisition protocol without retraining. Uses label maps + generative model to create unlimited training data.
    • Breakthrough: Eliminated the domain gap problem in brain segmentation. Now the default in FreeSurfer 7.2+.
    • 🔗 Download

  • SAMSEG: Longitudinal Whole-Brain and White Matter Lesion Segmentation — Cerri et al., 2023

    • Key idea: Sequence-adaptive segmentation using a generative model that simultaneously estimates tissue contrasts and segments. Longitudinal extension models within-subject changes over time.
    • 🔗 Download

  • Deep Learning Enables Automatic Detection and Segmentation — General review of DL-based medical image segmentation approaches.

Key Concepts

  • Atlas-based methods (FreeSurfer, SAMSEG) use probabilistic templates but are slow; deep learning methods (FastSurfer, SynthSeg) are fast but need training data.
  • SynthSeg's domain randomization paradigm — training on synthetic data to achieve domain-invariant models — is a breakthrough applicable beyond neuroimaging.
  • Longitudinal segmentation (SAMSEG) is critical for tracking disease progression over time.

Brain Tumor Detection & Segmentation

Papers

  • Brain Tumor Segmentation with Deep Neural Networks — Review

    • Key idea: Survey of deep learning approaches for brain tumor segmentation, covering U-Net variants, multi-scale architectures, and the BraTS challenge.
    • 🔗 Download (search arXiv, BraTS challenge papers)

  • A Deep Learning Approach for Brain Tumor Classification

    • Key idea: CNN-based classification of brain tumors from MRI — distinguishing gliomas, meningiomas, and pituitary tumors.
    • 🔗 Download

  • Automatic Brain Tumor Detection and Segmentation Using U-Net — Dong et al., 2017

    • Key idea: Comprehensive review of methods from traditional (region growing, atlas-based) to deep learning (FCN, U-Net, attention) for tumor detection.
    • 🔗 Download

  • Raidionics: Brain and Lesion Segmentation — Bouget et al., 2021/2023

    • Key idea: Open-source framework for preoperative and postoperative brain tumor segmentation with automatic standardized reporting. Designed for clinical integration.
    • Significance: Bridges the gap between research models and clinical deployment.
    • 🔗 Download

Key Concepts

  • U-Net architecture (encoder-decoder with skip connections) remains the dominant paradigm for medical image segmentation.
  • Multi-modal MRI (T1, T1ce, T2, FLAIR) provides complementary information — tumors appear differently on each contrast.
  • Clinical integration (Raidionics) requires not just accuracy but also fast inference, standardized reporting, and robustness to varied scan quality.

Stroke & Ischemic Lesion Segmentation

Papers

  • Artificial Intelligence for Diagnosing Acute Stroke: A 25-Year Retrospective — Wang et al., 2024

    • Key idea: Reviews AI for acute stroke imaging — detection of large vessel occlusion, ischemic core estimation, penumbra identification, and hemorrhage detection. Covers the clinical pipeline from symptom onset to AI-assisted diagnosis and treatment decision.
    • 🔗 Download

  • Artificial Intelligence in Stroke Imaging: A Comprehensive Review — Koska & Selver, 2023

    • Key idea: Covers AI applications across the stroke imaging pipeline — automated ASPECTS scoring, perfusion analysis, vessel occlusion detection, and outcome prediction.
    • 🔗 Download

  • Automatic Brain Ischemic Stroke Segmentation with Deep Learning: A Review — Abbasi et al., 2023

    • Key idea: Focused review on DL architectures for ischemic lesion segmentation — 2D vs 3D networks, attention mechanisms, and multi-scale processing.
    • 🔗 Download

  • Automatic Segmentation in Acute Ischemic Stroke: Prognosis on Stroke Volumes and Outcome — Wong et al., 2022

    • Key idea: Used automated lesion segmentation to predict clinical outcomes — larger infarct volumes correlate with worse functional status at 90 days.
    • 🔗 Download (search NeuroImage: Clinical 2022)

  • Group-Derived and Individual Disconnection in Stroke Prognosis — Bey et al., 2025

    • Key idea: Modeled stroke outcomes using brain disconnection patterns (graph-based) rather than lesion volume alone, showing that network disruption better predicts functional impairment.
    • 🔗 Download

Key Concepts

  • Stroke is a time-critical emergency — AI can reduce image interpretation time from minutes to seconds.
  • Disconnection-based prognosis (Bey) is superior to simple lesion volume because the same-sized lesion in different locations disrupts different brain networks.
  • The field is transitioning from lesion segmentation alone to outcome prediction and treatment decision support.

Anomaly Detection in Brain Imaging

Papers

  • Unsupervised Brain Anomaly Detection and Segmentation with Transformers — Walter Hugo Lopez Pinaya et al., 2021

    • Key idea: Used variational autoencoders to learn normal brain appearance, then detected anomalies as reconstruction errors. No labeled pathological data needed for training.
    • 🔗 Download

  • AnoDDPM: Anomaly Detection with Denoising Diffusion Models — Wyatt et al., 2022

    • Key idea: Diffusion models denoise brain MRI — the difference between input and denoised output reveals anomalies that the model cannot explain as normal variation.
    • 🔗 Download

  • Unsupervised Lesion Detection in Brain CT Using Bayesian CNN — Pawlowski et al., 2018

    • Key idea: Combined Bayesian CNNs (MC Dropout) with VAE-based anomaly detection to detect lesions in CT scans. Uncertainty estimates help distinguish true anomalies from reconstruction artifacts.
    • 🔗 Download

Key Concepts

  • Unsupervised anomaly detection requires only healthy training data — learns "normal" then flags deviations. Crucial for rare pathologies.
  • Diffusion models (AnoDDPM) outperform VAEs for anomaly detection because they produce sharper reconstructions.
  • Combining anomaly scores with uncertainty estimates reduces false positives from reconstruction artifacts.

← Previous: Neuroimaging Pipelines | → Next: Normative Modeling