ABUS-ResMask-Net: ABUS Lesion Classification Network Design with ResMask Module and BI-RADS Content-Awareness...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Automated Breast Ultrasound (ABUS) systems are crucial for breast lesion diagnosis.
• Accurate classification of benign versus malignant breast lesions is essential for patient management.
• Deep learning models show promise in improving diagnostic performance in medical imaging.

Purpose of the Study:

• To develop and evaluate a novel 2.5D deep learning model, ABUS-ResMask-Net, for benign-malignant classification of breast lesions using ABUS images.
• To improve diagnostic performance by incorporating shape and margin features as auxiliary tasks.
• To compare the proposed model against existing state-of-the-art methods.

Main Methods:

• A retrospective study involving 387 breast lesions from 313 patients across two centers.
• Development of a 2.5D deep learning model (ABUS-ResMask-Net) using Swin Transformer V2-T backbone and a ResMask Fusion module.
• Integration of shape and margin classification as auxiliary tasks, aligned with BI-RADS categories.
• External validation on a held-out cohort from a different center.

Main Results:

• The ABUS-ResMask-Net achieved a lesion-level Area Under the Curve (AUC) of 0.91 (95% CI: 0.83-0.96).
• The model outperformed benchmark methods, including 3D Swin Transformer (AUC 0.76), Yang et al. (AUC 0.85), and BI-RADS-Net-v2 (AUC 0.87).
• Superior performance was demonstrated on the external testing cohort.

Conclusions:

• The proposed 2.5D ABUS-ResMask-Net effectively classifies benign and malignant breast lesions in ABUS images.
• The model's architecture, incorporating auxiliary tasks and the ResMask Fusion module, enhances diagnostic accuracy.
• ABUS-ResMask-Net shows significant potential for clinical application in breast lesion diagnosis.