Topology-Aware Hybrid Vision Transformer–EfficientNet Framework for Calibrated Breast Ultrasound Classification

Abstract

Breast cancer detection using ultrasound imaging has gained increasing attention because ultrasound is affordable, non-invasive, and useful for examining dense breast tissue where mammography may show reduced sensitivity. However, ultrasound-based diagnosis remains challenging due to speckle noise, low contrast, irregular lesion boundaries, imaging variability, and limited generalization of existing deep learning models. This study proposes a topology-aware hybrid deep learning framework for three-class breast ultrasound classification. The framework combines boundary-sensitive tokenization, EfficientNet–Vision Transformer feature fusion, radiomics-preserving augmentation, calibrated probabilistic inference, and self-distillation to classify ultrasound images as normal, benign, or malignant. OMTAT converts ultrasound images into topology-aware patch tokens while preserving lesion structure and boundary information; DiSCo-TS integrates EfficientNet-based local representations with Vision Transformer-based global context; CRPA introduces controlled image variations while constraining radiomic drift; FROCED calibrates predictive probabilities; and X-MESD regularizes learned representations through self-distillation. Experimental evaluation on the BUSI dataset reports 96.8% accuracy, 95.9% precision, 95.2% recall, 95.5% F1-score, and 0.98 ROC-AUC under the stated image-level split. The BUSI lesion masks are used only as anatomical reference information where applicable; segmentation is not treated as an evaluated output task in the present study. The current evidence therefore supports within-dataset breast ultrasound classification and calibrated decision support, while patient-independent and external generalization require dedicated validation.