BUS-M2AE: Multi-scale Masked Autoencoder for Breast Ultrasound Image Analysis

Area of Science:

• Medical imaging
• Artificial intelligence
• Oncology

Background:

• Masked AutoEncoder (MAE) shows promise in medical image analysis by reducing annotation needs.
• Existing MAE models struggle with ultrasound breast tumor images due to varying tumor sizes and morphologies.
• This limits their generalization for accurate breast cancer diagnosis.

Purpose of the Study:

• To develop a novel Masked AutoEncoder model tailored for ultrasound breast tumor images.
• To enhance the model's ability to generalize across diverse tumor scales and morphologies.
• To improve breast cancer classification and segmentation using ultrasound data.

Main Methods:

• Proposed Breast UltraSound Multi-scale Masked AutoEncoder (BUS-M2AE) model.
• Implemented multi-scale masking at both token and feature levels.
• Utilized vision transformers for adaptive feature perception.

Main Results:

• BUS-M2AE demonstrated superior performance in breast cancer classification and tumor segmentation.
• The multi-scale masking effectively addressed limitations of general MAE models with varying tumor sizes.
• Outperformed recent MAE variants and supervised learning methods.

Conclusions:

• BUS-M2AE offers an effective solution for analyzing diverse breast tumor morphologies in ultrasound images.
• The proposed multi-scale masking strategy enhances MAE's adaptability for medical imaging tasks.
• BUS-M2AE holds significant potential for improving automated breast cancer diagnosis.