Integrating machine learning with histopathological, immunohistochemical, and PET parameters in the study of invasive
Berkay Çağdaş1, Hüseyin Şan2, Kardelen Çağdaş3
1Department of Nuclear Medicine, Afyonkarahisar State Hospital, Afyonkarahisar, Türkiye.
Background:
This study investigates the relationship between histopathological (HP) features, immunohistochemical (IHC) markers, 18F- FDG PET/CT parameters, and machine learning algorithms in patients diagnosed with invasive breast carcinoma, no special type.
Methods:
384 patients were included in the study. 18F- FDG PET/CT images after diagnosis of invasive breast carcinoma, before treatment, were analyzed for the PET metabolic parameters retrospectively. PET/CT images of patients who met the inclusion criteria and were retrospectively analysed. Metabolic parameters including SUVmax, SUVmean, SUVpeak, MTV (Metabolic Tumour Volume), and TLG (Total Lesion Glycolysis) were calculated, considering SUVmax at 42% of the threshold value in the primary lesion for the PET/CT imaging. None of the parameters fit into the normal distribution. Therefore, Mann-Whitney-U, Kruskal-Wallis, and Spearman correlation tests were applied. The machine learning algorithms were also used to describe any connection between the HP and IHC properties of the tumors.
Results:
Significant correlations were observed between PET-derived parameters (SUVmax, SUVmean, SUVpeak, MTV, and TLG) and HP/IHC characteristics, including Nottingham grade, molecular subtypes, and Ki-67 proliferation index. Machine learning models (XGBoost, Random Forest, LightGBM, CatBoost) demonstrated moderate success in identifying high-risk pathological and phenotypic features, with SUVmean identified as the most influential predictor via SHAP analysis. The coefficient of variation revealed low to moderate variability across most performance metrics, with CV values predominantly between 0.01 and 0.09, indicating stable and reliable model outputs.
Conclusion:
These findings suggest that integrating 18 F-FDG PET/CT metabolic parameters and machine learning algorithms can enhance non-invasive risk stratification and guide personalized treatment strategies in breast cancer management.
