Artificial Intelligence Methods Applied to PET-Based Assessment of Ischemic Stroke

Introduction: 1. Review and summarize the literature regarding the application of artificial intelligence (AI) methods to positron emission tomography (PET) imaging in the context of ischemic stroke diagnosis and management. 2. Evaluate the potential accuracy and reliability of AI-assisted diagnosis and the potential for improved patient outcomes. 3. Propose recommendations for the future development and implementation of AI in PET imaging for stroke.

Methods: A comprehensive search of the literature was conducted using the following electronic databases: PubMed, Scopus, and Web of Science. Eligible studies for inclusion were peer-reviewed papers that applied AI methods to PET imaging for stroke diagnosis or management. Data were then extracted from eligible studies, including information regarding study design, sample size, AI methodologies utilized, limitations, and research findings. Information from the included studies was synthesized and analyzed to address the aforementioned objectives.

Results: Artificial intelligence (AI) and deep learning, when applied to PET imaging, have the potential to improve outcomes for stroke patients, especially those with acute ischemic stroke. One challenge in the treatment of this condition is reducing the "door-to-puncture" time for thrombolysis, a procedure used to break up blood clots.

Researchers have attempted to use deep convolutional neural networks (CNNs) to predict oxygen extraction fraction (OEF) maps from other PET and MR images, without the need for a [¹⁵O]-PET scan. Previous research has applied convolutional neural networks (CNNs) to learn structural MR images and arterial spin labeling maps (cerebral blood flow and cerebral blood volume) to predict oxygen extraction fraction (OEF) maps. OEF maps were estimated based on images acquired through inhaled 15O2, and cerebral blood flow was estimated from H₂¹⁵O PET images. The predicted OEF maps were similar to the actual maps, and the learned model had an intraclass correlation of 0.60, indicating that CNNs trained with PET and MR images can qualitatively predict OEF maps without the need for a dedicated separately acquired inhaled [¹⁵O]-PET scan. The literature suggests that processing time to obtain OEF mapping can be reduced, which could accelerate stroke treatment protocols.

After a stroke event, up to one-third of survivors report post-stroke dementia (PSD). Because stroke patients already have decreased metabolism in their affected brain region, it is often difficult to identify the degree to which the decreased metabolism would predict dementia as shown with [¹⁸F]-fluorodeoxyglucose (FDG)-PET.

Recently, a 3D CNN was utilized in the prediction of dementia after training with a dataset of FDG-PET scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The algorithm was then applied to a dataset of PET scans from patients with stroke. The model was able to differentiate Alzheimer's disease from normal controls in an independent stroke cohort. This suggests that the resulting FDG-PET cognitive signature could be used as an independent risk factor for dementia following stroke.

Conclusions: The application of AI and deep learning to PET imaging has the potential to improve outcomes for stroke patients and aid in the evaluation of stroke rehabilitation treatments. Deep learning algorithms can be used to restore missing PET imaging data and predict OEF maps from other PET and MR images, potentially reducing the need for time-intensive [¹⁵O]-PET imaging. In addition, CNN models trained on large datasets of FDG-PET have the potential to be used as a biomarker for predicting cognitive impairment in patients with both neurodegenerative disorders and cerebrovascular diseases. To fully understand the potential of AI in the diagnosis and treatment of stroke, further research should be conducted using prospective study designs and interventional approaches.