BUViTNet: Breast Ultrasound Detection via Vision Transformers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Convolutional neural networks (CNNs) have improved breast cancer detection using ultrasound images.
• Vision transformers (ViTs) excel in natural image analysis, surpassing CNNs with enhanced global information processing and skip connections.
• The efficacy of ViTs in breast ultrasound imaging remains unexplored.

Purpose of the Study:

• To investigate the effectiveness of Vision Transformers (ViTs) for breast ultrasound image analysis.
• To introduce BUViTNet, a novel ViT-based deep learning model for breast ultrasound detection.
• To evaluate BUViTNet's performance against existing methods in classifying breast ultrasound images.

Main Methods:

• Developed BUViTNet, a ViT-based model employing multistage transfer learning.
• Utilized ImageNet and cancer cell image datasets for initial transfer learning.
• Trained and evaluated the model on two public datasets: Mendeley and breast ultrasound images (BUSI).

Main Results:

• BUViTNet achieved perfect scores (AUC, MCC, kappa = 1 ± 0) on the Mendeley dataset.
• On the BUSI dataset, BUViTNet attained high scores: AUC of 0.968 ± 0.02, MCC of 0.961 ± 0.01, and kappa of 0.959 ± 0.02.
• BUViTNet significantly outperformed ViT trained from scratch, conventional ViT transfer learning, and CNN-based transfer learning (p < 0.01).

Conclusions:

• Vision transformers demonstrate significant potential for analyzing breast images and enhancing early breast cancer detection.
• BUViTNet offers improved diagnostic accuracy, suggesting its utility in clinical settings.
• Further research with diverse datasets and parameters can optimize transformer-based breast imaging analysis.