[A multiscale carotid plaque detection method based on two-stage analysis].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diagnostics

Context:

• Carotid artery plaques are crucial indicators of cardiovascular disease risk.
• Accurate plaque identification in ultrasound images is challenging due to variations in scale and appearance.
• Existing detection methods may lack the precision and speed required for clinical settings.

Purpose:

• To develop and evaluate a robust deep learning-based method for accurate multiscale carotid plaque detection in ultrasound images.
• To enhance the precision and efficiency of carotid plaque identification using a two-stage approach.
• To compare the proposed method against established object detection models.

Summary:

• A two-stage deep learning model, SM-YOLO, was developed for carotid plaque detection.
• The first stage uses a YOLOX_l network with multiscale strategies for candidate plaque generation.
• The second stage employs Histogram of Oriented Gradient (HOG) and Local Binary Pattern (LBP) features with a Support Vector Machine (SVM) classifier for refined detection.

Impact:

• SM-YOLO demonstrated superior performance with 90.96% accuracy and 92.70% AP, outperforming other models.
• The method achieves real-time detection capabilities, significantly faster than Faster R-CNN.
• This advancement offers potential for improved diagnostic accuracy and efficiency in carotid ultrasound analysis.