Automated Radiomics Analysis from Multi-Modal Image Segmentation for Predicting Triple Negative Breast Cancer.

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Artificial Intelligence in Medicine

Background:

• Triple-negative breast cancer (TNBC) presents diagnostic challenges.
• Accurate differentiation of TNBC from non-TNBC is crucial for effective treatment planning.
• Current diagnostic methods may lack the quantitative precision needed for optimal classification.

Purpose of the Study:

• To investigate the utility of quantitative radiomic features from PET/CT scans in distinguishing TNBC from non-TNBC.
• To develop and evaluate a computational pipeline combining deep learning for segmentation and machine learning for classification.
• To assess the potential of radiomics as a non-invasive tool for TNBC identification.

Main Methods:

• Retrospective analysis of PET/CT images from 217 breast cancer patients (57 TNBC, 160 non-TNBC).
• Automated tumor segmentation on PET images using a deep learning model, mapped to CT scans.
• Extraction of radiomic features from 3D tumor volumes and classification using machine learning with recursive feature elimination.
• Performance evaluation using F1-score, AUC, accuracy, sensitivity, and specificity via 5-fold cross-validation.

Main Results:

• The proposed radiomic approach achieved high performance metrics.
• Key metrics included an F1-score of 0.90 ± 0.02, accuracy of 0.86 ± 0.07, and AUC of 0.88 ± 0.04.
• Top-ranked radiomic features significantly contributed to the classification accuracy.

Conclusions:

• Quantitative radiomic features from PET/CT scans offer valuable prognostic insights for TNBC identification.
• Machine learning algorithms combined with automated PET/CT segmentation can accurately differentiate TNBC from non-TNBC.
• This image-based radiomic analysis presents a promising non-invasive tool to improve TNBC diagnosis and treatment strategies.