Automatic detection of anomalies in screening mammograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Breast cancer screening diagnostic performance is affected by disease probability.
• High probability of normal screening exams can lead to false negatives.
• A system to pre-sort mammograms by disease likelihood was investigated.

Purpose of the Study:

• To develop a classification system for pre-sorting screening mammograms into normal and suspicious categories.
• To investigate the use of wavelets for parsing image data and potentially removing confounding information.
• To assess the feasibility of improving diagnostic performance by adjusting prior probabilities.

Main Methods:

• Mammograms were processed using 2D discrete wavelet transforms to create multi-scale maps.
• Statistical features were extracted from wavelet maps and used as input for naïve Bayesian classifiers.
• The classifier network calculated the probability of abnormality in mammography images, tested on DDSM and MIAS databases.

Main Results:

• The developed classifier networks achieved 100% sensitivity and up to 79% specificity.
• This performance surpasses the mean sensitivity of unaided human experts.
• The system demonstrated potential utility in a clinical setting for abnormality detection.

Conclusions:

• Wavelet-derived features enabled highly sensitive classification across various pathologies, nearly eliminating Type II errors.
• Pre-sorting mammograms significantly altered the prior probability of disease in the tested database.
• The developed system shows promise for enhancing the accuracy and efficiency of breast cancer screening programs.