Purpose: This study aimed at predicting four important immunohistochemical biomarkers, including estrogen receptor, progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki67 (cell proliferation rate index)) in breast cancer using radiomic features derived from multicentric ¹⁸F-FDG PET/CT images.

Methods: Sixty-two patients with locally advanced breast cancer who underwent ¹⁸F-FDG PET/CT imaging before any treatment were included. Three different PET/CT scanner models were used to acquire the images. After tumor segmentation, radiomic features from PET and CT images were extracted using the Python PyRadiomics package. Fusion features were created at the feature level, including concatenation (Con) and averaging (Avg). Combat was applied for features harmonization. The area under the curve (AUC), sensitivity, and specificity were used to evaluate the performance of predictive models.

Results: Random Forest (RF) model in Con features with mean AUC of 0.69 ± 0.11, Support Vector Machine (SVC) model in radiomic features from CT with a mean AUC of 0.74 ± 0.02 were outstanding in predicting ER and PR, respectively. The best models for predicting HER2 were RF and SVC using CT images, with mean AUC of 0.72 ± 0.04 and 0.73 ± 0.03. Respectively. In addition, Ki67 was predicted on radiomic features derived from PET images by RF and SVC models with mean AUC of 0.8 ± 0.09 and 0.83 ± 0.03, respectively.

Conclusion: Machine learning classifiers based on PET, CT, and PET/CT radiomic features could be correlated with the immunohistochemical biomarkers in breast cancer.