Ensemble Learning for Breast Cancer Lesion Classification: A Pilot Validation Using Correlated Spectroscopic Imaging...

Area of Science:

• Medical Imaging
• Spectroscopy
• Machine Learning

Background:

• Accurate classification of breast masses is crucial for timely cancer diagnosis.
• Conventional imaging techniques have limitations in differentiating benign from malignant lesions.
• Advanced spectroscopic and diffusion imaging offer novel biomarkers for breast cancer assessment.

Purpose of the Study:

• To evaluate machine learning models for classifying malignant and benign breast masses.
• To assess the utility of features derived from 5D EP-COSI and DWI for breast mass classification.
• To identify key spectral and diffusion parameters predictive of malignancy.

Main Methods:

• Extracted 2D correlated spectroscopy spectra and DWI data from breast masses.
• Utilized metabolite/lipid ratios, water/fat fractions, and apparent diffusion coefficients (ADC) as features.
• Employed recursive feature elimination and sequential forward selection to identify significant features.
• Trained and evaluated individual and ensemble machine learning models, including XGBoost and GradientBoost.

Main Results:

• XGBoost and GradientBoost models achieved high performance (AUC > 93%, Accuracy > 85%) in classifying breast masses.
• Identified ADC, specific 2D spectral peaks, and cross-peaks as critical features for classification.
• Conventional biomarkers like choline, myo-inositol, and glycine were significant predictors.

Conclusions:

• Machine learning models, particularly XGBoost and GradientBoost, demonstrate strong potential for accurate breast mass classification.
• The integration of 5D EP-COSI derived spectral features significantly enhances classification performance.
• This approach shows promise for improving early breast cancer detection and diagnosis.