A multi-modal ensemble framework for breast cancer segmentation and classification using genetic U-Net and HBoost

Breast cancer remains a leading cause of mortality among women worldwide, and early, accurate diagnosis is essential to improving treatment outcomes. Traditional segmentation and classification models often fail in ultrasound imaging due to noisy inputs, imbalanced class distributions, and computational inefficiencies. To address these challenges, we propose M2G-HBoost, a multi-modal ensemble framework explicitly designed for robustness in noisy and unbalanced data conditions. The framework integrates cosine similarity–based Graph Convolutional Network (GCN) augmentation to enrich feature diversity by modeling global topological and attribute relationships, M2GCNet for joint spatial and channel-wise dependency modeling, Genetic U-Net optimized via evolutionary algorithms for parameter-efficient high-accuracy segmentation, and HBoost, a heterogeneous boosting ensemble, for resilient classification. This modular design was chosen to employ complementary strengths: GCN augmentation for data diversity, M2GCNet for rich feature extraction, Genetic U-Net for segmentation precision with low complexity, and HBoost for classification stability under imbalance. On the BUSI dataset, M2G-HBoost achieved a Dice score of 93.58 % and IoU of 91.81 %, outperforming CBAM-RIUnet (76.25 % Dice, 80.73 % IoU), PCA-UNet (80.47 %, 84.12 %), EDCNN (85.99 %, 86.24 %), ELRL-E (87.14 %, 88.67 %), and SegEIR-Net (91.07 %, 89.49 %). In classification, the model reached 96.45 % (benign), 98.53 % (malignant), and 98.87 % (normal), exceeding AdaBoost, XGBoost, and Gradient Boosting. These results demonstrate the superiority and clinical applicability of the proposed method, offering a robust, accurate, and efficient solution for breast cancer segmentation and classification in ultrasound imaging.