Early Breast Cancer Prediction Using Deep Multiscale Fusion and Nomadic People Optimization

Early diagnosis of breast cancer is crucial for successful prognosis, early treatment, and enhanced survival rates. In this paper, a single multimodal predictive analytics approach is presented where the Deep Multiscale Fusion Neural Network (DMFNN) is optimized by the Nomadic People Optimizer (NPO), a new metaheuristic motivated by the adaptive migration and survival tactics of nomadic tribes. The NPO-based DMFNN differs from traditional optimizers like PSO, GA, or DE; NPO converges faster and more stably, strikes a good balance between exploration and exploitation, and prevents premature convergence to local optima. Such ability enables more accurate and efficient parameter tuning and enhances classification performance, robustness, and generalization for various imaging modalities. The framework proposed incorporates preprocessing, feature extraction, and classification into an end-to-end pipeline for processing the heterogeneous nature of three datasets: CBIS-DDSM (mammograms), Breast Histopathology Images (BHI), and Breast Cancer Patients’ MRIs (BCP MRI). Preprocessing is done using a standardized Multilayer Adaptive Guided Side Window Box Filter to maintain a uniform quality of images. To extract robust features, Fast Discrete Curvelet Transform with Wrapping (FDCT-WRP) is used to extract multiscale and directional features essential for breast cancer examination. The optimized DMFNN-NPO model is subsequently employed to make precise predictions. Deployment using Python demonstrates that the framework obtains great performance with the best performance on the BCP MRI dataset: accuracy of 97.12%, F1-score of 95.23%, precision of 93.78%, recall of 94.89%, and ROC AUC of 98.12%. The performances prove the efficacy of fusing deep multimodal learning and NPO-optimized optimization for early breast cancer prediction, providing strong integration potential into clinical decision support systems.