An Optimized Framework for Breast Cancer Classification Using Machine Learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Early breast cancer detection significantly improves survival rates.
• Increasing ultrasound image volume strains radiologist capacity, leading to potential misclassification and high false-positive rates.
• Accurate classification of breast lesions is crucial for timely and effective treatment.

Purpose of the Study:

• To develop and evaluate a computer-aided diagnosis (CAD) system for automated breast lesion classification.
• To optimize machine learning algorithms for improved accuracy in distinguishing malignant from benign breast tumors.
• To address the limitations in radiologist analysis of large volumes of breast ultrasound data.

Main Methods:

• Utilized a machine learning approach, selecting 13 key features from 185 available for model training.
• Implemented five distinct machine learning classifiers for binary classification of tumors.
• Employed Bayesian optimization with a tree-structured Parzen estimator for algorithm optimization and 10-fold cross-validation.

Main Results:

• The LightGBM classifier demonstrated superior performance compared to the other four models.
• Achieved high performance metrics: 99.86% accuracy, 100.0% precision, 99.60% recall, and 99.80% F1-score.
• The optimized CAD system effectively classified malignant versus benign breast tumors.

Conclusions:

• The proposed computer-aided diagnosis system significantly enhances the accuracy of breast lesion classification.
• The LightGBM classifier, optimized through Bayesian methods, offers a robust solution for automated analysis of breast ultrasound images.
• This technology holds potential to alleviate radiologist workload and improve diagnostic outcomes in breast cancer screening.