Deep learning-based interpretation of basal/acetazolamide brain perfusion SPECT leveraging unstructured reading...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neurology

Background:

• Basal/acetazolamide brain perfusion single-photon emission computed tomography (SPECT) is crucial for assessing functional hemodynamics in carotid artery stenosis.
• Interpreting SPECT findings often relies on unstructured text reports, posing challenges for consistent analysis.

Purpose of the Study:

• To develop a deep learning model to support the interpretation of brain perfusion SPECT.
• To leverage unstructured text reports for automated extraction of structured diagnostic information.

Main Methods:

• A dataset of 7345 basal/acetazolamide brain perfusion SPECT images and reports was retrospectively collected.
• A Long Short-Term Memory (LSTM) network was trained to extract structured labels (basal perfusion, vascular reserve) from text reports.
• A 3D Convolutional Neural Network (CNN) model was developed using extracted data to interpret SPECT images, with performance evaluated using AUC.

Main Results:

• The LSTM model achieved high AUCs (1.00 for basal perfusion, 0.99 for vascular reserve) for structured label extraction.
• The CNN model demonstrated strong performance in identifying abnormal perfusion (AUC 0.83 for basal perfusion, 0.89 for vascular reserve).
• CNN model output showed significant improvement post-revascularization and correlated with clinical outcomes.

Conclusions:

• A deep learning model was successfully developed to convert unstructured SPECT text reports into structured labels, supporting interpretation.
• This AI-powered system can identify perfusion abnormalities and provide quantitative scores, benefiting patient management, particularly those undergoing revascularization.