Improved segmentation of cardiac structures in echocardiograms for diastolic function evaluation.

Area of Science:

• Medical imaging and artificial intelligence
• Cardiovascular diagnostics
• Echocardiography analysis

Background:

• Noninvasive assessment of diastolic dysfunction relies on complex echocardiographic measurements with high interobserver variability.
• Automated semantic segmentation of cardiac structures (left atrium, left ventricle, mitral valve annulus) can capture temporal changes for improved analysis.

Purpose of the Study:

• To enhance the accuracy of segmenting the left atrium, left ventricle, and mitral valve annulus in echocardiographic images.
• To utilize temporal segmentation features for more reliable identification of diastolic dysfunction.

Main Methods:

• Development of Diff-TransUNet, a novel segmentation model featuring a noise-robust Differential Transformer module.
• Evaluation on private, CAMUS, and EchoNet-Dynamic datasets using Dice coefficient (Dice), Intersection-over-Union (IoU), and Hausdorff Distance (HD95) metrics.
• Statistical analysis with Benjamini-Hochberg FDR correction and Cohen's d effect size to assess performance improvements.

Main Results:

• Diff-TransUNet demonstrated superior performance across all datasets, achieving high Dice, IoU, and low HD95 values.
• Significant improvements over state-of-the-art models were observed, with Dice improvements ranging from 0.42%-4.96% (p < 0.05).
• Segmentation features extracted by the model achieved 88.95% accuracy in identifying diastolic dysfunction.

Conclusions:

• The Diff-TransUNet model significantly enhances ultrasound image segmentation accuracy.
• Segmented features of the left ventricle, left atrium, and mitral annulus effectively aid in diastolic dysfunction identification.