Benchmarking Supervised and Self-Supervised Learning Methods in a Large Ultrasound Muti-Task Images Dataset.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning in ultrasound (US) imaging is hindered by limited datasets and narrow task types.
• Foundational models require datasets that accurately reflect US imaging's unique characteristics.

Purpose of the Study:

• Introduce US-MTD120K, a large-scale, multi-task ultrasound dataset.
• Provide a rich basis for training and evaluating deep learning models in US imaging.
• Benchmark state-of-the-art methods in supervised and self-supervised learning for US image analysis.

Main Methods:

• Collected and curated 120,354 real-world 2D ultrasound images.
• Developed a dataset covering standard plane recognition and diagnostic tasks.
• Benchmarked 27 supervised and self-supervised learning methods.

Main Results:

• Identified that constraining global feature computation can be a viable approach for US image analysis in supervised learning.
• Proposed MoCo-US, an improved self-supervised learning strategy that reduces reliance on pretext task design.
• MoCo-US achieves competitive performance and enhances other self-supervised learning methods.

Conclusions:

• US-MTD120K dataset facilitates advancements in deep learning for ultrasound imaging.
• The proposed MoCo-US strategy offers an effective approach for self-supervised learning in medical imaging.
• Future research can leverage this dataset and methods for improved US image analysis.