Development of a Clinically Applicable Deep Learning System Based on Sparse Training Data to Accurately Detect Acute...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Non-enhanced head computed tomography (CT) is crucial for diagnosing acute intracranial hemorrhage in head trauma and stroke.
• Accurate and timely detection of intracranial hemorrhage is vital for clinical decision-making.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm (DeepCT) for detecting acute intracranial hemorrhage on non-enhanced head CT.
• To assess the clinical applicability and universal reliability of the DeepCT algorithm.

Main Methods:

• A deep learning model based on U-Net and ResNet architectures was developed using PyTorch.
• The model was trained on 1,815 CT image sets and validated on independent datasets from multiple centers, including US and Taiwan.
• Performance was evaluated using accuracy metrics across validation and test datasets.

Main Results:

• The DeepCT algorithm demonstrated high accuracy, with overall accuracy ranging from 0.9343 to 0.9820 across datasets.
• The algorithm effectively identified acute intracranial hemorrhage in non-enhanced head CT studies.
• Evaluation on US and Taiwan datasets confirmed the algorithm's universal reliability.

Conclusions:

• DeepCT is an effective deep learning tool for detecting acute intracranial hemorrhage on non-enhanced head CT.
• The algorithm can facilitate hyperacute triage, reduce reporting times, and enhance radiologist interpretation accuracy.
• The findings support the universal reliability of DeepCT for diagnosing intracranial hemorrhage in diverse clinical settings.