Abstract:
Pathological complete response (pCR) following neoadjuvant chemotherapy is a key prognostic
indicator in HR+/HER2+ (triple-positive) breast cancer. Early non-invasive prediction
of treatment response could help optimize therapeutic strategies and minimize
exposure to toxic treatments. This study aimed to predict pCR using radiomics features
extracted from apparent diffusion coefficient (ADC) maps derived from diffusion-weighted
MRI (DWI) trace images acquired at the early treatment stage (T0) from The Cancer
Imaging Archive (TCIA). The dataset included 215 ADC maps from 53 HR+/HER2+
patients, comprising 78 pCR and 137 non-pCR cases. ADC maps were generated by combining
DWI trace images with b-values of 0 and 800 s/mm2. The dataset was split into
70% training and 30% testing sets. Tumor regions of interest were manually delineated,
and first-order statistical, textural, and morphological radiomics features were extracted.
Correlation-based feature selection was applied to the training set, removing highly correlated
features with correlation >0.9. Class imbalance was addressed using the SMOTE
technique to upsample the minority class while minimizing overfitting. Several supervised
machine learning classifiers were evaluated. Random Forest achieved the highest test accuracy
(67.69%), followed by Decision Tree (61.54%), LightGBM (61.54%), Logistic Regression
(56.92%), Support Vector Machine (56.92%), XGBoost (58.46%), and Gradient
Boosting (47.69%). Five-fold cross-validation of Random Forest yielded 64.65% accuracy,
with an area under the ROC curve of 0.54. The optimal Random Forest parameters were:
n_estimators = 800, criterion = ‘entropy’, max_depth = 5, min_samples_split = 10,
min_samples_leaf = 8, max_features = 0.05, bootstrap = True, random_state = 42.
Limitations of this study include its retrospective multicenter design and small sample
size. These findings should be validated in larger cohorts. Nevertheless, this study provides
useful insights for predicting pCR in clinical practice, offering potential to guide
personalized treatment decisions in HR+/HER2+ breast cancer.
