Improving Computer-aided Detection for Digital Breast Tomosynthesis by Incorporating Temporal Change

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Digital breast tomosynthesis (DBT) is a key imaging modality for breast cancer screening.
• Existing deep learning frameworks for DBT lesion detection can be further improved.
• Temporal information, such as lesion changes over time, is valuable for accurate diagnosis.

Purpose of the Study:

• To develop and evaluate PriorNet, a deep learning algorithm utilizing temporal information to enhance DBT cancer detection.
• To refine an existing framework by incorporating growth information for improved malignancy probability assessment.

Main Methods:

• Retrospective analysis of DBT screening examinations from multiple institutions (2016-2020).
• Development of PriorNet as a cascaded deep learning module building upon a single-view detection and ipsilateral view matching framework.
• Training and validation on data from seven sites, with external testing on an eighth site; performance evaluated using localization ROC curves.

Main Results:

• PriorNet achieved a higher AUC (0.931) on the validation set compared to baseline models (0.892 and 0.915).
• On the external test set, PriorNet demonstrated superior performance with an AUC of 0.896 versus baselines (0.846 and 0.865).
• PriorNet showed significantly higher partial AUC in the high sensitivity range (0.9-1.0), indicating improved detection of true positives.

Conclusions:

• PriorNet effectively leverages temporal information to significantly improve the performance of DBT cancer detection frameworks.
• The algorithm demonstrates enhanced accuracy in distinguishing malignant lesions by considering changes over time.
• This deep learning approach offers a promising advancement for computer-aided detection in breast cancer screening.