Deep Transfer Learning for Automatic Prediction of Hemorrhagic Stroke on CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Intracerebral hemorrhage (ICH) is a critical form of stroke requiring prompt diagnosis.
• Manual analysis of noncontrast-computed tomography (NCCT) brain images for ICH is labor-intensive and time-consuming.
• Accurate and rapid diagnosis is crucial for effective patient management.

Purpose of the Study:

• To develop and evaluate an automated deep transfer learning model for detecting intracranial hemorrhage on NCCT brain images.
• To improve the efficiency and accuracy of hemorrhagic stroke diagnosis.
• To assess the model's potential as a clinical decision support tool.

Main Methods:

• A deep transfer learning model combining ResNet-50 and a dense layer was proposed.
• The model was trained and evaluated on 1164 NCCT brain images from 62 patients.
• The model classifies individual CT images as either hemorrhagic or normal.

Main Results:

• The proposed deep transfer learning model achieved high performance metrics.
• Achieved 99.6% accuracy, 99.7% specificity, and 99.4% sensitivity.
• Outperformed the standard ResNet-50 model in diagnosing intracranial hemorrhage.

Conclusions:

• The deep transfer learning model demonstrates significant advantages for automated hemorrhagic stroke diagnosis.
• The model shows potential as a valuable clinical decision support tool for radiologists.
• Automated analysis can expedite the diagnosis of intracranial hemorrhage, improving patient outcomes.