A Two-Phase Deep Learning Approach for Architectural Distortion Detection in Mammograms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women globally, necessitating early detection for improved outcomes.
• Architectural distortion (AD) is a subtle, early indicator of breast cancer on mammograms, challenging traditional detection methods.
• Automating AD segmentation and classification is crucial to reduce radiologist workload and enhance diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate an automated deep learning approach for segmenting and classifying architectural distortion (AD) in mammograms.
• To improve the accuracy and efficiency of early breast cancer detection by addressing the challenges of AD identification.
• To reduce the burden on radiologists by providing a reliable AI-assisted diagnostic tool.

Main Methods:

• A two-phase deep learning pipeline was developed, integrating U-Net++ for semantic segmentation and Mask R-CNN for instance segmentation.
• A ResNet-18 classification model was combined with Mask R-CNN to refine AD predictions and minimize false positives.
• Optimized loss functions, including smooth L1 and binary cross-entropy with Dice loss, were employed to enhance segmentation performance.

Main Results:

• The integrated deep learning approach achieved high performance metrics: segmentation accuracy of 0.852, classification accuracy of 0.915, and mean average precision (mAP) of 0.894.
• The system demonstrated a high sensitivity of 92.4% in detecting architectural distortion.
• The proposed method significantly improved segmentation metrics and reduced false positives compared to traditional approaches.

Conclusions:

• The developed deep learning pipeline offers a robust and accurate solution for automated AD segmentation and classification on mammograms.
• This AI-driven approach has the potential to significantly enhance breast cancer screening and diagnostic processes.
• Timely and accurate detection of AD through this technology can lead to improved patient outcomes via earlier treatment planning.