Towards Visual-Search Model Observers for Mass Detection in Breast Tomosynthesis

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Digital breast tomosynthesis (DBT) requires reliable image assessment for accurate cancer detection.
• Human observer models are crucial for evaluating the performance of imaging systems and interpretation tasks.
• Existing models like the channelized non-prewhitening (CNPW) observer have limitations in replicating complex search strategies.

Purpose of the Study:

• To develop and evaluate a novel visual-search observer model for DBT image analysis.
• To compare the performance of the visual-search model against the CNPW observer and human radiologists.
• To assess the model's ability to perform clinically realistic detection and localization tasks.

Main Methods:

• A two-phase visual-search observer model was implemented, combining holistic gradient template matching with CNPW-based analysis.
• Anthropomorphic breast phantoms with embedded spherical masses were used to generate simulated DBT projections.
• A localization receiver operating characteristic (LROC) study was conducted to compare observer performances.

Main Results:

• The visual-search observer model demonstrated performance comparable to human observers in terms of area under the LROC curve.
• The CNPW observer showed lower performance compared to both the visual-search model and human observers.
• The visual-search model successfully replicated aspects of trained radiologists' search patterns.

Conclusions:

• The developed visual-search observer model shows significant potential for realistic assessment of DBT images.
• The model's ability to achieve human-like performance suggests its utility in optimizing breast cancer screening technologies.
• Further research is necessary to enhance the algorithm's robustness for widespread clinical application.