Ipsilateral Lesion Detection Refinement for Tomosynthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computer-aided detection (CAD) for breast cancer screening has evolved significantly with deep learning.
• Multi-view analysis in CAD improves detection but often uses black-box models, hindering clinical trust and interpretability.
• Current multi-view CAD lacks explicit reasoning for lesion correspondence, complicating clinical decision-making.

Purpose of the Study:

• To develop a multi-view detection framework for digital breast tomosynthesis (DBT) that provides accurate cancer detection with transparent, step-by-step reasoning.
• To address the limitations of black-box deep-learning models in clinical settings by enabling analysis and fine-tuning.
• To introduce a novel approach for ipsilateral lesion matching to refine detection scores across multiple views.

Main Methods:

• Development of Ipsilateral-Matching-Refinement Networks (IMR-Net), a novel framework for DBT lesion detection.
• Integration of explicit ipsilateral lesion matching to adaptively refine single-view detection scores.
• Utilizing a commercial development DBT dataset comprising 24,675 volumetric views from 8,034 exams.

Main Results:

• The IMR-Net framework demonstrates improved performance in multi-view lesion detection.
• The explicit ipsilateral lesion matching refines detection scores, enhancing the reliability of the CAD system.
• Performance evaluation using location-based, case-level receiver operating characteristic (ROC) and free-response ROC (FROC) analysis.

Conclusions:

• IMR-Net offers a promising solution for interpretable multi-view breast cancer detection in DBT.
• The framework's ability to provide explicit reasoning facilitates clinical adoption and trust in AI-driven screening tools.
• This approach advances the development of transparent and effective CAD systems for breast cancer screening.