DGA3-Net: A parameter-efficient deep learning model for ASPECTS assessment for acute ischemic stroke using...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Early detection of acute ischemic stroke (AIS) using non-contrast computerized tomography (NCCT) is critical for timely treatment.
• Identifying subtle hypoattenuation and performing accurate Alberta Stroke Program Early CT Score (ASPECTS) assessments on NCCT are challenging and time-consuming for human experts.

Purpose of the Study:

• To develop and validate a novel convolutional neural network (CNN)-based model, DGA3-Net, for automated ASPECTS assessment by detecting early ischemic changes in NCCT scans.
• To improve the accuracy and efficiency of AIS diagnosis from NCCT.

Main Methods:

• DGA3-Net utilizes a parameter-efficient dihedral group CNN encoder, attention-guided slice aggregation, and an asymmetry-aware classifier.
• The model was trained and validated on retrospective NCCT datasets from suspected AIS patients (n=170 primary, n=90 external).
• Performance was evaluated against expert neuroradiologist assessments.

Main Results:

• DGA3-Net demonstrated superior performance compared to two expert neuroradiologists in both regional stroke identification (F1=0.69) and ASPECTS evaluation (Cohen's Kappa=0.70).
• Ablation studies confirmed the effectiveness of the model's design components.
• Visualization techniques showed that the model focused on clinically relevant imaging signs.

Conclusions:

• Deep learning, specifically the DGA3-Net model, shows significant potential for timely and accurate AIS diagnosis from NCCT scans.
• This AI-driven approach can enhance the quality of care for AIS patients by potentially reducing diagnostic time and improving assessment accuracy.