Tolerance to missing data using a likelihood ratio based classifier for computer-aided classification of breast...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Oncology

Background:

• Mammography is sensitive for breast tumor detection but has low specificity.
• Low specificity leads to a high rate (up to 70%) of unnecessary biopsies.
• Accurate differentiation between malignant and benign breast lesions is crucial.

Purpose of the Study:

• To develop a highly specific computer-aided diagnosis (CAD) algorithm.
• To improve the classification accuracy of mammographic masses.
• To reduce the number of unnecessary breast biopsies.

Main Methods:

• Developed a likelihood ratio-based classifier to handle missing data.
• Utilized a dataset of 671 biopsy-proven breast masses (245 malignant).
• Included 16 features from the BI-RADS lexicon and patient history, accommodating missing values.

Main Results:

• The classifier achieved 32% specificity at 100% sensitivity with 16 features, including those with missing values.
• Excluding cases with missing data (using only 7 features) decreased performance to 19% specificity at 100% sensitivity.
• The algorithm demonstrated the benefit of utilizing incomplete cases rather than discarding them.

Conclusions:

• The developed CAD algorithm shows commendable classification performance for mammographic masses at high sensitivity.
• This approach has the potential to spare benign breast lesions from unnecessary biopsies.
• Computer-aided diagnosis algorithms can effectively manage missing data in mammographic analysis.