Abstract:
Breast cancer grading plays a critical role in treatment planning and prognosis,
yet manual histopathological assessment is invasive, time-consuming, and subject
to inter-observer variability. Diffusion-Weighted Imaging (DWI) and Apparent
Diffusion Coefficient (ADC) maps provide quantitative information on tissue microstructure
and offer potential for non-invasive tumor characterization. This
study focused on developing and evaluating a machine learning (ML)-based radiomic
framework for classifying low-intermediate grade (Grade 1–2) and highgrade
(Grade 3) breast tumors from ADC-derived features. This study was carried
out using 846 labeled ADC maps from breast cancer patients who were radiologically
confirmed using the SBR (Scarff-Bloom-Richardson) grading system. Tumor
regions of interest were manually segmented, and first-order, textural, and
wavelet radiomic features were extracted. Feature redundancy and selection were
performed using Pearson/Spearman correlation and ANOVA F-test. Dataset imbalance
due to fewer low-intermediate-grade tumors was addressed using weighted
class learning, SMOTE, SMOTE-Tomek, and SMOTE-ENN. Logistic Regression,
Support Vector Machine (SVM), Random Forest, CatBoost, and XGBoost models
were trained using patient-wise splitting and cross-validation to prevent data
leakage. Performance was evaluated using accuracy, F1-score, and area under the
ROC curve (AUC). SVM combined with SMOTE-Tomek achieved the best performance,
yielding 68% accuracy and an AUC of 0.71, with improved sensitivity for
detecting high-grade tumors. These findings show that ADC-based radiomic features
capture clinically relevant markers for tumor aggressiveness. This proposed
framework provides a reproducible, non-invasive decision support tool that may
improve the diagnostic decision-making process. However, manual segmentation
may limit generalizability. Future studies should incorporate automated segmentation
and multimodal MRI features to enhance the robustness of the framework.
