Breast cancer is the leading cause of cancer-related mortality and morbidity among women worldwide. Early detection plays a crucial role in improving survival rates and BI-RADS classification is one of the effective ways of predicting breast cancer. However, BI-RADS Category 4 encompasses a broad spectrum of malignancy probabilities, ranging from over 2% to 95%. Due to the wide malignancy likelihood range and the ambiguous qualitative attributes of BI-RADS 4, patients are subjected to overdiagnosis and unnecessary procedures, such as biopsy, which entail a certain degree of physical trauma as well as financial strain. This study proposed a lightweight CNN where MobileNet serves as the backbone architecture, augmented with the Convolutional Block Attention Module (CBAM), resulting in the MobileNet-CBAM model. The model demonstrated good performance in discriminating BI-RADS 4 category malignant and benign cases in Contrast Enhanced Spectral Mammogram (CESM) with a prediction of 82%, 82% and 0.91 for accuracy, f1-score and roc-auc respectively. Additionally, for clinical friendliness, the model explanation was given using SHAP. Hence, the model presents potential utility in predicting breast cancer for lesions categorized as BI-RADS category 4 in breast imaging.