Automatic arterial input function selection in CT and MR perfusion datasets using deep convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate perfusion parameter imaging relies on identifying the arterial input function (AIF).
• Manual AIF extraction is time-consuming and subjective.
• Automated methods are needed for efficient and reliable AIF identification in clinical practice.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) for automatic arterial input function (AIF) identification.
• To assess the performance of the CNN method in computed tomography perfusion (CTP) and perfusion-weighted MRI (PWI) datasets.
• To compare the CNN-based AIF extraction with a traditional shape-based algorithm and manual selections.

Main Methods:

• A deep convolutional neural network (CNN) was trained and validated on 100 CTP and 100 PWI datasets from acute ischemic stroke patients.
• Manual AIFs and non-arterial tissue curves were used for training and validation.
• Performance was evaluated against expert manual AIFs and a shape-based algorithm using cross-correlation, shape features, and Dice similarity for hypoperfusion lesions.

Main Results:

• The CNN method demonstrated significantly higher cross-correlation values with manual AIFs compared to the shape-based method.
• Hypoperfusion lesions derived from CNN-extracted AIFs showed higher Dice similarity to manually derived lesions.
• Shape features of CNN-derived AIFs closely matched manual AIFs, with slightly higher peak heights in PWI.

Conclusions:

• Deep convolutional neural networks (CNNs) provide a viable and accurate method for automatic AIF extraction from CTP and PWI data.
• The proposed CNN approach enhances the reliability and efficiency of perfusion imaging analysis in stroke patients.
• This automated method holds promise for improving clinical decision-making in acute ischemic stroke.