Local curvature analysis for classifying breast tumors: Preliminary analysis in dedicated breast CT

Area of Science:

• Radiology
• Medical Imaging
• Computational Pathology

Background:

• Accurate breast tumor classification is crucial for effective treatment.
• Novel image features can enhance the diagnostic capabilities of medical imaging.
• Current methods may benefit from incorporating advanced feature extraction techniques.

Purpose of the Study:

• To evaluate the effectiveness of local curvature measures as new image features for breast tumor classification.
• To compare the performance of curvature measures against traditional image features.
• To assess the potential of curvature measures in simplifying classification models.

Main Methods:

• 119 breast lesions from 104 CT scans were analyzed.
• Volumetric segmentation and surface extraction were performed.
• Total, mean, and Gaussian curvatures were computed and normalized for classification.
• Traditional image features were extracted and compared using logistic regression and cross-validation.

Main Results:

• Normalized total curvature (CT) achieved an AUC of 0.74.
• A combination of five traditional features yielded an AUC of 0.83.
• CT demonstrated potential to replace other features, enabling simpler classifiers.
• A classifier combining RGI and CT achieved an AUC of 0.81.

Conclusions:

• Normalized curvature measures provide valuable information for breast tumor classification.
• Incorporating curvature measures can lead to reduced feature sets and potentially more robust classifiers.
• This approach may enhance the reliability of diagnostic models across different datasets.