Breast cancer is the most prevalent cancer among women and predicting pathologic complete response (pCR) after anti-cancer treatment is crucial for patient prognosis and treatment customization. Deep learning has shown promise in medical imaging diagnosis, particularly when utilizing multiple imaging modalities to enhance accuracy. This study presents a model that predicts pCR in breast cancer patients using dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) and apparent diffusion coefficient (ADC) maps. Radiomics features are established hand-crafted features of the tumor region and thus could be useful in medical image analysis. Our approach extracts features from both DCE MRI and ADC using an encoder with a self-attention mechanism, leveraging radiomics to guide feature extraction from tumor-related regions. Our experimental results demonstrate the superior performance of our model in predicting pCR compared to other baseline methods.