ichseg

ichseg performs automated segmentation of intracerebral hemorrhage (ICH) in X-ray computed tomography (CT) scans to quantify hemorrhage location and volume for clinical research and treatment planning.

Key Features:

• Automated Segmentation: Fully automated segmentation of intracerebral hemorrhage in CT scans using machine learning.
• Random Forest Algorithm: A random forest classifier predicts voxel-level probabilities of hemorrhage presence in CT scans.
• Imaging Predictors from MISTIE: Imaging predictors are extracted from baseline CT scans of patients enrolled in the Minimally Invasive Surgery plus rt-PA in ICH Evacuation (MISTIE) trial.
• Multiple Model Development: Four models were developed: logistic regression, LASSO-penalized logistic regression, generalized additive model (GAM), and a random forest classifier.
• Model Comparison and Performance: Model performance was evaluated against manual expert segmentations using the Dice Similarity Index (median DSI for the random forest = 0.899), with the random forest outperforming logistic regression, LASSO, and GAM.
• Statistical Validation: Statistical tests with Bonferroni correction were used to confirm significant performance differences between the random forest and other models.
• Volume Correlation: High correlation between manual and predicted hemorrhage volumes was reported (correlation coefficient 0.93, 95% CI: 0.9–0.95).
• Voxel Selection and Thresholding: A first-pass voxel selection based on quantiles of derived imaging predictors is used, and predicted probabilities are thresholded to generate binary segmentations.

Scientific Applications:

• Clinical Research: Quantification of ICH location and volume for clinical research and outcome analyses.
• Multi-site Clinical Trials: Provides consistent and reproducible segmentation and volume measurements across imaging centers, supporting trials such as MISTIE.
• Stroke Management and Treatment Planning: Enables objective volumetric assessment to inform clinical decision-making and treatment planning in intracerebral hemorrhage.

Methodology:

Imaging predictors are extracted from baseline CT scans from the MISTIE trial; a first-pass voxel selection based on quantiles of derived predictors is applied; models (logistic regression, LASSO, GAM, random forest) are trained (training using 10 randomly selected scans with remaining scans for validation); the random forest predicts voxel-level probabilities which are thresholded to produce binary segmentations that are compared to manual segmentations using the Dice Similarity Index and statistical tests with Bonferroni correction; volume correlation is reported (r=0.93, 95% CI: 0.9–0.95).