Innovative pathomics approach: integrating radiomic and CNN features for breast cancer classification using histopathological images

This study aims to improve the accuracy and reliability of grading invasive ductal carcinoma (IDC) into four grades (1–4) by developing an integrated framework combining radiomic features and deep learning-derived features. The proposed framework seeks to address current diagnostic challenges by leveraging both traditional radiomics and advanced deep learning approaches to enhance classification performance and clinical applicability. The dataset comprised 925 whole-slide images (WSIs) divided into over 356,000 image patches categorized into four grades. Radiomic features were extracted using the HistomicsTK platform and grouped into five categories, including shape, texture, and intensity-based features. Deep features were obtained from fine-tuned DenseNet169 and EfficientNet models. Two distinct approaches were employed: in the first, deep features were extracted and combined with radiomic features as input to machine learning classifiers (XGBoost, LightGBM, CatBoost, Random Forest). In the second, an end-to-end framework directly utilized CNNs for multiclass classification. Advanced feature selection techniques such as LASSO, ANOVA, and RFE were used to identify the most predictive features. Evaluation metrics included accuracy, sensitivity, and AUC-ROC, and ensemble models were implemented to further enhance performance. The integration of radiomic and deep learning features outperformed all standalone approaches, achieving a test AUC-ROC of 95.9%, accuracy of 94.1%, and sensitivity of 93.9% with RFE and XGBoost. Ensemble models combining DenseNet169 and EfficientNet further improved results, achieving a test AUC-ROC of 95.2% and accuracy of 94.44%. SHAP analysis identified key features such as Intensity.Min, Shape.EquivalentDiameter, and Haralick.Entropy.Mean as critical for classification. The t-SNE visualization confirmed the discriminative power of selected features. This study demonstrates the efficacy of integrating radiomic and deep learning features for IDC grading, offering a robust and clinically interpretable framework. Future work could incorporate molecular data to further enhance predictive accuracy.