Automated Detection of the Black Hole Sign for Patients with Intracerebral Hemorrhage Using Self-Supervised Learning

Area of Science:

• Artificial Intelligence in Radiology
• Medical Image Analysis
• Stroke Imaging

Background:

• Intracerebral Hemorrhage (ICH) is a severe stroke type.
• Hematoma Expansion (HE) on CT scans predicts poor outcomes.
• The Black Hole Sign (BHS) on CT indicates HE risk.

Purpose of the Study:

• Develop an automated method for BHS identification.
• Leverage self-supervised learning (SSL) for improved accuracy.
• Facilitate early patient selection for interventions.

Main Methods:

• A ResNet-50 encoder was pre-trained on 1.7M+ unlabeled head CT images.
• SSL techniques were employed for BHS identification.
• The model was fine-tuned and evaluated on diverse datasets for AUC and F1 scores.

Main Results:

• The SSL framework achieved strong BHS identification performance.
• Scan-level AUC scores ranged from 0.75-0.89, with F1 scores of 0.60-0.70.
• The model demonstrated generalizability on external datasets, with AUC up to 0.85.

Conclusions:

• SSL integration offers a powerful approach for medical image classification.
• The automated BHS identification reduces manual annotation needs.
• This framework shows promise for predicting HE in ICH patients, especially with limited data.