Do Deep Learning Algorithms Accurately Segment Intracerebral Hemorrhages on Noncontrast Computed Tomography? A...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Stroke is a leading global health concern, with intracerebral hemorrhage (ICH) volume critical for treatment and prognosis.
• Conventional ICH volume calculation methods (e.g., ABC/2) can be inaccurate due to shape assumptions.
• Deep learning (DL) shows promise in enhancing noncontrast computed tomography (CT) for ICH volume estimation.

Purpose of the Study:

• To systematically review and meta-analyze the precision of deep learning algorithms for delineating ICH on noncontrast CT.
• To assess the performance of DL models in ICH segmentation compared to traditional methods.

Main Methods:

• A systematic review and meta-analysis of studies from 2000-2023, adhering to PRISMA guidelines.
• Inclusion of 28 studies, primarily retrospective cohorts, focusing on convolutional neural network architectures (e.g., U-Net).
• Performance evaluated using Dice Similarity Coefficient (DSC); risk of bias assessed with the Prediction Model Risk of Bias Assessment Tool.

Main Results:

• Meta-analysis of 14 studies yielded a combined DSC of 0.85 (95% CI, 0.82-0.88), indicating high accuracy.
• DL model performance was consistent across methodologies but higher for spontaneous ICH compared to other types.
• U-Net variants were the predominant deep learning architectures employed.

Conclusions:

• Deep learning models demonstrate high efficacy in segmenting ICH on noncontrast CT, offering potential clinical neuroimaging advancements.
• Challenges persist in accurately segmenting smaller hemorrhages, necessitating further research.
• DL has the potential to reduce healthcare professional workload and improve patient care in stroke management.