Automatic breast mass detection in mammograms using density of wavelet coefficients and a patch-based CNN.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning in Radiology

Background:

• Accurate mass detection in digital mammograms is crucial for early breast cancer diagnosis.
• Existing methods face challenges with false positives, impacting diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate a two-stage automated method for accurate mass detection in full-field digital mammograms.
• To enhance diagnostic accuracy by reducing false positives while maintaining high sensitivity.

Main Methods:

• A two-stage approach involving suspicious region localization using density of wavelet coefficients (DWC-QLS) and false-positive reduction via a CNN classifier.
• Exploration of transfer learning strategies, including fine-tuning the InceptionV3 model, for improved classification performance.

Main Results:

• The proposed method achieved a True Positive Rate (TPR) of 0.98 at 1.43 False Positives Per Image (FPI) on the INbreast database.
• Achieved 100% sensitivity in mass location detection with an average of 5.4 false positives per image.

Conclusions:

• The developed method successfully detects and classifies suspicious regions in digital mammograms.
• The proposed approach demonstrates superior True Positive Rate and False Positive Index performance compared to state-of-the-art methods.