Lesion Asymmetry Screening Assisted Global Awareness Multi-View Network for Mammogram Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammography is crucial for early breast cancer detection.
• Current deep learning models often analyze images independently, limiting diagnostic performance.
• Integrating multi-view mammographic data is essential for improved accuracy.

Purpose of the Study:

• To develop an end-to-end deep learning framework for patient-level breast cancer diagnosis.
• To address limitations of image-level analysis by incorporating multi-view interactions.
• To improve diagnostic performance and interpretability in mammography screening.

Main Methods:

• Proposed a novel framework: lesion asymmetry screening assisted global awareness multi-view network (LAS-GAM).
• LAS-GAM operates at the patient level, processing four mammographic views (CC and MLO).
• Employs a global module for comprehensive patient assessment and a lesion screening module for targeted feature extraction.

Main Results:

• LAS-GAM achieved Area Under the Curve (AUC) scores of 0.817 on the DDSM dataset and 0.894 on an in-house dataset.
• The patient-level approach demonstrated superior diagnostic performance compared to image-level models.
• Training with patient-level labels significantly reduced data annotation costs.

Conclusions:

• LAS-GAM effectively integrates multi-view mammographic information for improved breast cancer diagnosis.
• The framework simulates radiologist workflows, enhancing both performance and interpretability.
• This patient-centric deep learning approach offers a promising advancement in mammography-based screening.