Cascaded Deep Learning-Based Model for Classification and Segmentation of Plaques from Carotid Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Carotid plaque classification and burden quantification are vital for stroke risk assessment.
• Current methods may lack synergy between plaque classification and segmentation.

Purpose of the Study:

• To develop an integrated framework for joint carotid plaque classification and segmentation.
• To enhance the performance of both tasks by sharing plaque location information.
• To improve stroke risk stratification using 2D ultrasound images.

Main Methods:

• A cascaded framework integrating Masked-ResNet-DS (ResNet-based classifier) and MedSAM (medical Segment Anything Model).
• Utilizing ground truth boundaries for region-specific feature pooling in the classifier during training.
• Implementing a two-iteration strategy with a class activation map (CAM) for focused pooling and segmentation guidance at inference.
• Supervising CAM using Dice loss against segmentation ground truth for accurate localization.

Main Results:

• Masked-ResNet-DS achieved a mean F1-score of 96.7% for plaque classification, outperforming other methods by at least 3.2%.
• CAM-guided MedSAM achieved a Dice Similarity Coefficient (DSC) of 86.6%, surpassing U-Net and nnU-Net.
• CAM prompts improved MedSAM's DSC by 2.2%, demonstrating the benefit of shared location information.

Conclusions:

• The proposed framework effectively leverages shared plaque location information to enhance both classification and segmentation tasks.
• This integrated approach offers a more accurate tool for carotid plaque analysis and stroke risk stratification.
• The synergy between classification and segmentation leads to superior performance compared to standalone methods.