Robust fine-grained echocardiographic view classification with supervised contrastive learning.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate echocardiographic view classification is crucial for automated cardiac analysis.
• Clinical practice faces challenges due to heterogeneous acquisitions and inter-observer variability.
• Existing studies often use limited view sets, reducing clinical relevance.

Purpose of the Study:

• Introduce TTE47, the first public benchmark with 47 clinically meaningful echocardiographic views.
• Enable rigorous quantification of inter-observer agreement and reproducible evaluation.
• Develop a robust framework for fine-grained echo view classification.

Main Methods:

• Propose a novel supervised contrastive learning framework with a tailored loss function.
• Utilize a dataset (TTE47) with views annotated independently by three experts.
• Introduce clustering-based metrics (Detection Rate, Label Recovery Precision) for semantic coherence analysis.

Main Results:

• The proposed method outperforms cross-entropy and standard contrastive baselines on TTE47.
• Achieved leading performance on TTE47 and surpassed prior work on TMED-2.
• Learned feature space aligns with anatomical structure, showing resilience to annotation variability and human-level disagreement.

Conclusions:

• The TTE47 dataset and proposed method establish a scalable, clinically relevant framework for fine-grained echo view classification.
• Contrastive pretraining can standardize interpretation, mitigate subjectivity, and enhance AI-assisted echocardiography reliability.
• The approach demonstrates robustness comparable to human inter-observer variability.