Stacked Bidirectional Convolutional LSTMs for Deriving 3D Non-Contrast CT From Spatiotemporal 4D CT

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Neurology

Background:

• Acute stroke imaging relies on non-contrast CT (NCCT) and CT perfusion (CTP).
• Deriving NCCT from CTP could streamline the imaging workup, reducing patient radiation dose and time.
• Current methods may not fully leverage the spatiotemporal information within CTP data.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting NCCT from CTP images.
• To assess the potential of this method to simplify acute stroke imaging protocols.
• To demonstrate the efficacy of a convolutional LSTM (C-LSTM) network for this task.

Main Methods:

• A stacked bidirectional convolutional LSTM (C-LSTM) network was designed to predict 3D NCCT volumes from 4D CTP data.
• The C-LSTM network was trained on 17 CTP-NCCT image pairs.
• Performance was quantitatively evaluated on a separate cohort of 16 cases and validated on larger independent datasets.

Main Results:

• The C-LSTM network significantly outperformed baseline and other convolutional neural network methods in predicting NCCT from CTP.
• The method demonstrated good scalability and performance on independent datasets with diverse pathologies.
• Successful derivation of NCCT from CTP was achieved, indicating potential for workflow simplification.

Conclusions:

• The C-LSTM network presents a promising deep learning approach for deriving NCCT from CTP images in acute stroke.
• This technique can potentially reduce imaging workup time and radiation dose.
• The C-LSTM network shows potential as a generalizable tool for analyzing high-dimensional spatiotemporal medical imaging data.