Multi domain, multi-scale diagnostic modeling for histopathological breast cancer classifications

Abstract

Breast cancer detection through histopathological imaging remains challenging due to complex tissue morphology, observer variability, and subtle differences between invasive and pre-invasive lesions. Conventional computer-aided diagnostic systems often rely on single-domain feature extraction, restricting multi-scale representation and clinical interpretability. To overcome these limitations, we propose a verified diagnostic framework integrating five analytical components for efficient and explainable breast cancer classification. The adaptive multi-level histopathological feature selection using cross-domain mutual information maximization (AMFSCDMIM) extracts highly informative morphological and frequency features with minimal redundancy. The deep hierarchical hybrid morphological– frequency encoding network (DH-HMFEN) refines spatial–spectral representations, while the multi-scale morphological attention classification network (MS-MACNet) applies adaptive attention across tissue structures for improved discrimination. The adaptive ensemble validation for breast cancer classification (AEV-BCC) calibrates confidence levels for enhanced reliability, and the comparative analytical performance validation with interpretability integrated metrics (CAPV-IIM) quantitatively evaluates model explainability using expert annotations. Experimental results on benchmark datasets achieve 96% accuracy, 0.98 area under the receiver operating characteristic curve (AUROC), and a 0.88 interpretability alignment score, outperforming existing methods. The proposed confidence-calibrated, multi-domain, and multi-scale framework enhances diagnostic precision and clinical trust in histopathology-based breast cancer detection.