Fusion of Radiomics and Deep Learning Features for Enhanced Prediction of Neoadjuvant Chemotherapy Response in Breast Cancer

Abstract

INTRODUCTION: Pathological complete response (pCR) following neoadjuvant chemotherapy (NAC) is a validated surrogate endpoint for long-term survival in breast cancer patients. However, conventional biomarkers exhibit limited predictive accuracy, with approximately 60-80% of patients failing to achieve pCR. Dynamic contrast-enhanced MRI (DCE-MRI) provides high-resolution information on tumor vascularization and heterogeneity, but prior radiomics models have predominantly relied on single-feature paradigms, which may not fully capture complex tumor phenotypes.

METHODS: We developed a multimodal deep-learning radiomics (DLR) pipeline using the publicly available ACRIN 6657/I-SPY1 dataset (n=163). After rigorous preprocessing (bias-field correction, isotropic resampling, Z-score normalization), we extracted a comprehensive set of 1,702 standardized radiomics features compliant with the Image Biomarker Standardization Initiative (IBSI), which quantitatively capture tumor morphology, texture, and intensity patterns. Additionally, 8,576 deep learning features were derived from five convolutional neural networks (ResNet50, DenseNet-169, InceptionV3, InceptionResNetV2, EfficientNetB0), enabling the model to learn complex, data-driven representations beyond human-defined features. The fusion of these complementary feature types provides a more holistic characterization of tumor phenotype, significantly enhancing predictive performance compared to single-modality approaches. A two-stage feature-selection strategy utilizing univariate analysis and the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was applied, followed by linear signature construction. Ten classifiers were evaluated under stratified cross-validation and independent testing.

RESULTS: The fusion of handcrafted radiomics and deep learning features significantly enhanced predictive performance. The best-performing model, a multilayer perceptron (MLP), achieved an area under the receiver operating characteristic curve (AUC) of 0.98 on the independent test set, with an accuracy of 95.92%, sensitivity of 92.86%, and specificity of 97.14%. Logistic regression also demonstrated strong performance (AUC = 0.980). Decision curve analysis confirmed the clinical utility of all models across a wide range of threshold probabilities.

CONCLUSIONS: The integration of radiomics and deep learning features within a machine learning framework provides a robust, non-invasive tool for predicting pCR to NAC in breast cancer. This multimodal approach outperforms single-modality models and offers potential for clinical translation to personalize treatment strategies and avoid ineffective chemotherapy. Further multi-center validation is warranted to confirm its generalizability.