Assessing Ischemic Stroke with Convolutional Image Features in Carotid Color Doppler

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Stroke is a major global cause of death and disability.
• Timely and precise diagnosis of acute stroke is crucial for effective treatment and prognosis.
• Carotid color Doppler (CCD) imaging is a key diagnostic tool.

Purpose of the Study:

• To develop and evaluate an automated system using convolutional neural networks (CNNs) for identifying acute ischemic stroke lesions in CCD images.
• To assess the diagnostic performance of different CNN models and explore the utility of individual color channels.

Main Methods:

• A retrospective analysis of 1032 CCD images from 106 acute ischemic stroke patients and 79 controls.
• Development and evaluation of various CNN models, including transfer learning approaches, using 10-fold cross-validation.
• Comparison of diagnostic performance across different CNN architectures and color channels, with neuroradiologist consensus as the gold standard.

Main Results:

• The CNN model trained from scratch (AlexNet) achieved high accuracy (91.67%), sensitivity (93.33%), specificity (90.20%), and AUC (0.9432).
• Other transfer learning models showed accuracies ranging from 77.69% to 83.94%.
• The green color channel demonstrated the best performance among individual channels (Accuracy: 87.50%, AUC: 0.9507).

Conclusions:

• The proposed CNN-based computer-aided diagnosis system effectively utilizes automatic feature extraction from CCD images for ischemic stroke prediction.
• This automated approach shows significant potential for clinical application in assisting stroke diagnosis.