Breast Tumor Characterization Using [¹⁸F]FDG-PET/CT Imaging Combined with Data Preprocessing and Radiomics

Area of Science:

• Radiology and Nuclear Medicine
• Oncology
• Medical Imaging Analysis

Background:

• Breast cancer diagnosis and subtyping remain critical challenges in oncology.
• Current methods using standard uptake value (SUV) lesion classification have limitations.
• Novel approaches integrating advanced imaging features are needed for improved accuracy.

Purpose of the Study:

• To evaluate ensemble learning holomic models for breast cancer detection and subtype classification using [18F]FDG-PET/CT.
• To assess the impact of data pre-processing algorithms on model performance.
• To compare machine learning (ML) models with conventional SUV-based analysis.

Main Methods:

• A cohort of 170 patients with 173 breast tumors underwent [18F]FDG-PET/CT imaging.
• Radiomic features were extracted following IBSI guidelines with optimized extraction.
• Ensemble learning with five supervised ML algorithms and 100-fold Monte Carlo cross-validation was employed.
• Data pre-processing included outlier detection, feature selection, and class imbalance correction.

Main Results:

• The cancer detection model achieved 80% sensitivity, 78% specificity, and 0.81 AUC.
• The triple-negative tumor identification model showed 85% sensitivity, 78% specificity, and 0.82 AUC.
• Models for individual receptor status and luminal A/B subtypes had lower performance (0.46-0.68 AUC).
• SUVmax yielded 0.76 AUC for cancer detection and 0.70 AUC for triple-negative subtype prediction.

Conclusions:

• Ensemble learning models with [18F]FDG-PET/CT and advanced pre-processing enhance breast cancer diagnosis.
• These models show promise for ML-based prediction of the aggressive triple-negative breast cancer subtype.
• Further development could improve prediction accuracy for other subtypes and receptor statuses.