MultiASNet: Multimodal Label Noise Robust Framework for the Classification of Aortic Stenosis in Echocardiography

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Aortic stenosis (AS) is a common heart valve disorder requiring early detection, yet diagnosis is challenging in routine practice.
• Echocardiography with Doppler is the standard but requires specialists, limiting accessibility.
• Point-of-care ultrasound (POCUS) is accessible but lacks Doppler analysis for AS screening.

Purpose of the Study:

• To develop a multimodal machine learning framework, MultiASNet, for enhanced AS screening using POCUS.
• To integrate 2D B-mode POCUS videos with structured echocardiography report data, including Doppler parameters.
• To enable reliable AS screening in non-specialist settings by overcoming limitations of current POCUS.

Main Methods:

• MultiASNet employs contrastive learning to align video and tabular report features from the same patient.
• Cross-attention in a transformer network addresses misalignment between single reports and multiple video views.
• The model uses sample selection to mitigate label noise from observer variability and integrates structured data during training only.

Main Results:

• MultiASNet achieved 93.0% balanced accuracy for AS detection on a private dataset and 83.9% on the TMED-2 dataset.
• For AS severity classification, balanced accuracy reached 80.4% (private) and 59.4% (public TMED-2).
• The framework demonstrated improved accuracy by counteracting label noise.

Conclusions:

• MultiASNet facilitates reliable AS screening in non-specialist settings by leveraging POCUS B-mode videos and report data.
• The model effectively bridges the diagnostic gap left by the absence of Doppler data in basic POCUS.
• This AI-driven approach reduces noise-related errors and enhances the interpretability of POCUS for AS diagnosis.