Deep learning-based correction for time truncation in cerebral computed tomography perfusion.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Cerebral computed tomography perfusion (CTP) imaging is crucial for diagnosing ischemic stroke.
• Clinical CTP acquisition often suffers from time truncation, leading to incomplete contrast bolus data.
• Incomplete data can compromise the accuracy of perfusion parameter quantification.

Purpose of the Study:

• To develop and evaluate a deep learning approach for predicting missing CTP image frames.
• To compare different prediction strategies for handling time truncation in CTP.
• To assess the impact of the prediction method on image quality and clinical parameters.

Main Methods:

• A three-dimensional convolutional neural network (CNN) was trained to predict the last 10 frames of CTP series.
• Seventy-two CTP scans were used for training and testing the CNN models.
• Three prediction strategies were evaluated: single-shot, recursive multi-step, and direct-recursive hybrid prediction.

Main Results:

• Single-shot prediction, predicting all missing frames simultaneously, yielded superior image quality metrics compared to recursive methods.
• The single-shot approach demonstrated the highest correlation (r=0.990) and lowest variance in bolus shape analysis.
• Perfusion parameters derived from single-shot predictions showed the smallest absolute differences from ground truth.

Conclusions:

• The proposed CNN-based approach effectively predicts missing CTP frames, mitigating time truncation issues.
• Single-shot prediction is the optimal strategy for accurate CTP data reconstruction and analysis.
• This method holds potential for minimizing errors and enhancing the quantification of ischemic stroke.