Deep learning-based interpretation of cerebrovascular reserve on basal/acetazolamide stress brain perfusion SPECT

Abstract

Early and accurate detection of cerebrovascular disease is important for its mortality and brain injury. Basal/acetazolamide stress brain perfusion single photon emission computed tomography (SPECT) is a functional diagnostic imaging tool to detect cerebral perfusion decrease and cerebrovascular reserve. The visual interpretation of brain perfusion SPECT image is a standard practice in the clinical setting, often resulting interobserver variability and inconsistence of diagnosis. In this study, we applied Long Short-Term Memory (LSTM) network and 3D convolutional neural network (CNN) model for the deep learning-based interpretation of the text report and image of basal/acetazolamide stress brain perfusion SPECT. LSTM network was successfully trained to classify the text report of each image regarding its hemodynamic abnormality. The LSTM model-predicted results were used for the label of a cerebrovascular reserve decrease on basal/acetazolamide stress brain perfusion SPECT images to train 3D CNN model. Our designed 3D CNN model was trainable but did not show outstanding performance to detect the cerebrovascular reserve decrease on basal/acetazolamide stress brain perfusion SPECT images. Our results suggest that 3D CNN is a trainable model on basal/acetazolamide stress brain perfusion SPECT in the detection of a cerebrovascular reserve decrease using text report prediction of LSTM as a ground truth label. Additional image preprocessing steps with advanced network architecture are required to improve the performance of our deep-learning based interpretation system in future study.