Optimizing Automated Hematoma Expansion Classification from Baseline and Follow-Up Head Computed Tomography.

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Hematoma expansion (HE) is a key predictor of poor outcomes in intracerebral hemorrhage (ICH).
• Accurate HE assessment requires manual segmentation of hematomas on CT scans, which is time-consuming for large datasets.
• Automated segmentation methods can be hampered by cumulative errors, impacting HE classification accuracy.

Purpose of the Study:

• To develop and validate a computational pipeline for automated hematoma segmentation and HE classification in ICH.
• To integrate a deep-learning model to estimate probabilities of false HE classifications, thereby optimizing expert review.
• To reduce the manual workload and improve the efficiency of HE annotation in large-scale ICH research.

Main Methods:

• A tandem deep-learning classification model was combined with automated hematoma segmentation.
• Three multicentric cohorts (n=2261) were used for training, internal testing, and external validation.
• Ground truth binary HE annotations were generated for volumes ≥3, ≥6, ≥9, and ≥12.5 mL.

Main Results:

• The pipeline achieved efficient HE annotation, with a 95% sensitivity threshold excluding 47-88% of predictions from expert review.
• Less than 2% false-negative misclassification rates were observed in both internal and external validation cohorts.
• The strategy effectively minimized the misclassification rate while reducing the burden of expert review.

Conclusions:

• The developed pipeline offers a time-efficient and optimizable method for ground truth HE classification in large ICH datasets.
• Combining automated segmentation with deep learning probability measures enhances the reliability of HE assessment.
• This approach significantly reduces the manual expert review required for large-scale ICH studies, facilitating research on HE.