Core Concepts
Unsupervised method DDSB enhances phase detection accuracy in echocardiography.
Abstract
Standalone Note here
Introduction:
Accurate identification of End-Diastolic (ED) and End-Systolic (ES) frames is crucial for cardiac function assessment through echocardiography.
Traditional methods face limitations like data dependency, expert annotations, and lack of robustness.
Early Techniques:
Manual selection or simplistic criteria failed to capture heart dynamics accurately.
QRS complex onset and T wave's end were used but lacked practicality in emergency scenarios.
Attempts at quantifying similarity between ED and ES frames required manual selection.
Deep Learning Advancements:
Classification-based and regression-based models evolved significantly.
Recurrent Neural Networks (RNNs) integrated with CNNs showed promise.
Various models like TempReg-Net, DenseNet, GRU were explored for optimal results.
Proposed Method - DDSB:
Unsupervised and training-free approach for phase detection.
Utilizes distance-based strategy to enhance model robustness.
Achieves comparable performance on Echo-dynamic and CAMUS datasets.
Dataset and Metrics:
Adjustments made to CAMUS and Echo-dynamic datasets for experiments.
Evaluation based on mean absolute error (MAE) in frames for ED/ES detection.
Results and Discussion:
Comparative analysis with state-of-the-art methods on CAMUS and Echo-dynamic datasets.
Cross-dataset evaluation showcasing generalization capabilities of DDSB.
Ablation Study:
Effects of k directions on model performance stability.
Impact of change threshold α on method effectiveness.
Conclusion:
DDSB offers accurate phase detection without the need for extensive resources or training data.