A Novel Self-Supervised Learning-Based Method for Dynamic CT Brain Perfusion Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Dynamic CT perfusion imaging provides quantitative brain blood flow metrics (CBF, CBV, MTT).
• Low-dose CT protocols increase image noise, degrading perfusion map quality and reliability.
• Existing methods like SVD and TV struggle with noise-induced inaccuracies.

Purpose of the Study:

• To propose and evaluate a deep learning model (CNN-BiLSTM with attention) for self-supervised estimation of the impulse residue function (IRF) from noisy dynamic CT data.
• To compute brain perfusion parameters using the predicted IRF.
• To compare the proposed method against traditional techniques (SVD, TV) using simulated and real data.

Main Methods:

• Development of a convolutional neural network (CNN) and bi-directional long short-term memory (BiLSTM) model with an attention mechanism.
• Self-supervised learning to derive the impulse residue function (IRF) from dynamic CT images.
• Quantitative evaluation using simulated data and qualitative/quantitative comparison with singular value decomposition (SVD) and tensor total-variation (TV) on real patient data.

Main Results:

• Simulated data showed superior parameter estimation accuracy with the proposed deep learning method compared to SVD and TV.
• Real data analysis revealed visually similar perfusion maps but significant differences in parameter values.
• The deep learning approach uniquely identified longer mean transit time (MTT) in suspected infarct core regions.

Conclusions:

• The proposed deep learning method effectively yields reliable brain perfusion maps from noisy dynamic CT images.
• This AI-driven approach offers improved accuracy and potential for better clinical assessment of cerebrovascular conditions.
• The self-supervised IRF estimation overcomes limitations of traditional methods in low-dose CT perfusion imaging.