Learning from unbalanced data: a cascade-based approach for detecting clustered microcalcifications.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Detection

Background:

• Expert radiologists face challenges in identifying abnormalities in diagnostic images due to the prevalence of normal tissue.
• Designing Computer-Aided Detection (CADe) systems is complex because imbalanced training data (predominance of normal samples) hinders abnormality recognition.

Purpose of the Study:

• To present a novel approach for computer-aided detection that tackles the class imbalance problem.
• To develop and evaluate a system (CasCADe) for automated detection of clustered microcalcifications (μCs) using this novel approach.

Main Methods:

• A cascade of boosting classifiers is employed, with each node trained using a learning algorithm based on ranking, not classification error.
• This approach specifically addresses the severely unbalanced classification problem inherent in detecting clustered microcalcifications (μCs).
• The system was evaluated on a dataset of 1599 full-field digital mammograms from 560 cases.

Main Results:

• CasCADe demonstrated competitive performance when compared to the Hologic R2CAD ImageChecker, a commercial CADe system.
• At 90% lesion sensitivity on malignant cases, CasCADe achieved 0.13 false positives per image (FPpi) versus R2CAD's 0.21 FPpi (p=0.09).
• At 0.21 FPpi, CasCADe detected 93% of true lesions compared to R2CAD's 90% (p=0.11).

Conclusions:

• The proposed CasCADe system effectively handles class imbalance in medical image analysis.
• CasCADe shows potential to compete with high-end commercial Computer-Aided Detection systems for microcalcification detection.
• The ranking-based learning algorithm in boosting classifiers offers a viable solution for imbalanced datasets in CADe.