Attention-Based Deep Learning System for Classification of Breast Lesions-Multimodal, Weakly Supervised Approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women, necessitating improved early detection methods.
• Screening mammography is crucial for early diagnosis, but automated systems can enhance its effectiveness.
• Deep learning artificial intelligence offers potential for developing advanced automated diagnostic tools.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated breast cancer detection using mammography.
• To address data scarcity challenges in training AI models for medical image analysis.
• To improve the accuracy and reliability of early breast cancer diagnosis through artificial intelligence.

Main Methods:

• A Clustering-constrained Attention Multiple Instance Learning (CLAM) classifier was employed for training under data scarcity.
• Feature extractors (ResNet, EfficientNet) pre-trained on ImageNet were utilized.
• Multimodal-view classification using both CC and MLO mammography images was performed.

Main Results:

• The CLAM model achieved an AUC-ROC of 0.896 ± 0.017 on a private dataset and 0.848 ± 0.015 on the Chinese Mammography Database.
• High performance metrics including F1-score (81.8 ± 3.2), accuracy (81.6 ± 3.2), precision (82.4 ± 3.3), and recall (81.6 ± 3.2) were reported.
• Attentional maps provided insights into image features driving predictions, enhancing model explainability.

Conclusions:

• The proposed CLAM-based deep learning approach effectively facilitates early breast cancer diagnosis from screening mammography.
• The model demonstrates superior performance compared to many existing studies, offering improved accuracy and some explainability.
• This AI-driven method holds promise for enhancing the efficiency and reliability of breast cancer screening programs.