Masked Modeling-Based Ultrasound Image Classification via Self-Supervised Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning methods dominate ultrasound data analysis but require extensive annotated datasets.
• Speckle noise and artifacts in ultrasound images create difficult classification challenges.
• Existing methods struggle with the scarcity of labeled ultrasound data.

Purpose of the Study:

• To develop a self-supervised pre-training method for ultrasound data analysis.
• To address the need for large annotated datasets in deep learning for ultrasound.
• To improve ultrasound image classification by handling hard examples and artifacts.

Main Methods:

• A novel pre-training method based on mask modeling for ultrasound data.
• Investigation of three masking strategies: random, vertical, and horizontal masking.
• Implementation of a hard sample mining strategy to manage difficult ultrasound images.

Main Results:

• The proposed method successfully extracts representative features from unlabeled ultrasound datasets.
• The approach demonstrates superior performance in ultrasound image classification compared to state-of-the-art methods.
• Effective feature extraction is achieved even with the presence of hard examples and imaging artifacts.

Conclusions:

• Self-supervised mask modeling is a viable and effective approach for ultrasound data pre-training.
• The method reduces reliance on human annotation, enabling the use of abundant unlabeled ultrasound data.
• The proposed strategy enhances the robustness and accuracy of ultrasound image classification models.