A completely automated CAD system for mass detection in a large mammographic database

Area of Science:

• Medical Imaging
• Computer-Aided Detection (CAD)
• Radiology

Background:

• Mass localization is vital for computer-aided detection (CAD) systems in mammography.
• Accurate detection of suspicious regions in mammograms aids in early cancer diagnosis.

Purpose of the Study:

• To present a fully automated classification system for detecting masses in digitized mammographic images.
• To develop a CAD tool that enhances the accuracy and efficiency of mammographic analysis.

Main Methods:

• Image segmentation using an iterative dynamical threshold algorithm to identify regions of interest (ROIs).
• ROI characterization via textural features from the gray tone spatial dependence matrix (GTSDM), ensuring invariance to monotonic transformations.
• ROI classification using a neural network trained with radiologist diagnoses.

Main Results:

• The CAD system was evaluated on 3369 mammographic images (2307 negative, 1062 positive).
• Receiver operating characteristic (ROC) analysis yielded an area under the curve (Az) of 0.783 +/- 0.008 for ROI-based classification.
• The system demonstrated 80% mass sensitivity with 4.23 false positives per image when compared to radiologist-drawn boundaries.

Conclusions:

• The developed automated CAD system shows promising performance in mass detection in mammograms.
• The use of invariant textural features allows for robust analysis across different imaging settings.
• This system can serve as a valuable tool for radiologists, potentially improving diagnostic accuracy and workflow.