Topological modeling and classification of mammographic microcalcification clusters

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Microcalcification clusters are key indicators of breast cancer.
• Classifying microcalcifications as malignant or benign is challenging and time-consuming for radiologists.
• Existing methods often focus on individual microcalcification morphology or global cluster statistics.

Purpose of the Study:

• To propose a novel method for classifying microcalcification clusters in mammograms.
• To utilize topology and connectivity analysis for improved classification accuracy.
• To develop a method that outperforms current state-of-the-art approaches.

Main Methods:

• Analysis of microcalcification topology/connectivity within clusters using multiscale morphology.
• Generation of microcalcification graphs to represent topological structure at different scales.
• Extraction of graph theoretical features for classification using k-nearest-neighbors classifiers.

Main Results:

• High classification accuracies (up to 96%) achieved on multiple datasets (MIAS, DDSM, full-field digital).
• Excellent ROC results with an area under the curve up to 0.96.
• Demonstrated superior performance compared to related publications through direct comparison.

Conclusions:

• The proposed topology modeling approach significantly outperforms existing methods for microcalcification cluster analysis.
• Topology modeling offers improved classification accuracy for microcalcifications.
• Topological measures provide insights that can be linked to clinical understanding of breast cancer.