Hierarchical Fuzzy System Integrated with Deep Learning for Robust and Interpretable Classification of Breast Malignancies Using Radiomics Features from Ultrasound Imaging

Breast cancer poses a global health risk and requires precision and accessibility in diagnostic measures. Ultrasound imaging is vital for breast lesion identification due to its safety, cost-effectiveness, and real-time capabilities. This paper presents a new fuzzy system architecture that utilizes ultrasound-based radiomics features to classify breast cancers. In order to ensure uniformity and consistency in shape-based characteristics limited to tumors, we calculate parameters such as elongation, compactness, spherical disproportion, and volumetrics following IBSI recommendations. We employ a hierarchical fuzzy system tree to handle high-dimensional data space and to identify the most discriminative characteristics. The selected features are incorporated into a modular fuzzy logic design that promotes transparency and maintains an auditable decision history according to clinical interpretability. Our framework enables the more accurate classification of breast cancer while addressing the beliefs and values prevalent in clinical applications. Tested on an independent set of data, the model achieved high accuracy of 99.60%, with low overfitting and strong generalization. To enhance its generalizability, we validated it on an internal dataset, attaining a sensitivity of 93.65%, a specificity of 99.24%, an AUC of 0.996, and an 18% reduction in unnecessary biopsies, as demonstrated through decision curve analysis, demonstrating substantial clinical utility across various settings. The findings confirm the system’s ability to identify intricate radiomic patterns linked to cancer. Due to its computing efficiency, it may be executed in real time during routine screening. The proposed radiomics-based fuzzy classification framework may offer a clinically beneficial approach for differentiating benign from malignant breast lesions. Explainability is enhanced with user-friendly artifacts for clinicians, including ranking IF-THEN rules and counterfactuals, all of which were validated in usability trials that demonstrated increased trust among radiologists compared to other technologies. Enhanced differentiation in the classification of various lesion types will decrease unnecessary biopsies. This approach integrates radiomics features with transparent and interpretable fuzzy logic to deliver enhanced predictors and a comprehensible framework for users, including physicians, to facilitate decision-making. This approach advances precision medicine standards through the early detection of lesions using more specific and systematic diagnostic instruments.