A New Semi-supervised Learning Benchmark for Classifying View and Diagnosing Aortic Stenosis from Echocardiograms.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cardiology

Background:

• Semi-supervised learning (SSL) advances have potential in clinical settings but require validation.
• Timely diagnosis of critical heart conditions like aortic stenosis is crucial.
• Cardiac ultrasound (echocardiogram) interpretation involves complex image analysis.

Purpose of the Study:

• To develop a benchmark dataset for evaluating SSL in echocardiogram interpretation.
• To assess SSL for cardiac view and disease severity classification.
• To improve patient-level diagnosis prediction using multi-task learning.

Main Methods:

• Developed a benchmark dataset for echocardiogram analysis.
• Applied state-of-the-art MixMatch SSL for view and severity classification.
• Implemented novel methods for patient-level diagnosis prediction, prioritizing clinically-relevant views and knowledge transfer.

Main Results:

• MixMatch SSL improved accuracy on view and severity classification tasks using limited labeled data.
• New patient-level methods outperformed baseline approaches by prioritizing relevant image views.
• Knowledge transfer from view classification enhanced patient-level diagnosis prediction.

Conclusions:

• SSL, particularly MixMatch, shows significant potential for improving echocardiogram interpretation.
• Prioritizing clinically-relevant views and multi-task learning are key for accurate patient-level diagnosis.
• The released dataset and framework will drive further advancements in clinical AI.