To develop and compare deep learning (DL) and conventional machine learning (ML) models based on carotid plaque ultrasound images, identifying the optimal clinically applicable algorithm for precise plaque assessment and risk prediction.
Key Findings:
ResNet-50 achieved the highest AUC of 0.982, with accuracy of 0.925, sensitivity of 0.964, and specificity of 0.897.
Logistic regression (LR) outperformed support vector machine (SVM) but without statistical significance (AUC: 0.885 vs. 0.861, p = 0.554).
ResNet-50 showed a 9.7% improvement in AUC over the best traditional ML model.
Interpretation:
The study demonstrates that the DL model ResNet-50 significantly outperforms traditional ML models in predicting stroke risk in patients with carotid plaques, indicating its potential as a valuable clinical diagnostic tool.
Limitations:
Study is retrospective and conducted at a single center, which may limit generalizability and applicability to broader populations.
The sample size, while substantial, may not encompass all variations of carotid plaque characteristics.
Conclusion:
The ultrasound image-based deep learning model ResNet-50 is validated as a superior tool for predicting stroke risk in patients with carotid plaques, suggesting its clinical applicability.
