A Multimodal Breast Cancer Diagnosis Approach Based on EfficientNet and Hybrid Feature Selection

Breast cancer ranks among the top causes of mortality in women globally. Traditional breast cancer diagnosis primarily relies on the prior experience of clinicians, which introduces a significant level of subjectivity to the process. Machine learning, as a technique that learns patterns from data and makes predictions, offers a more objective analysis and shows great potential in clinical diagnosis of breast cancer. However, most existing machine learning methods focus on a single modality, which limits the full utilization of data. Additionally, the black-box nature of end-to-end approaches leads to a lack of interpretability, introducing a degree of decision-making risk. Furthermore, irrelevant and redundant features inherent in the data further affect model performance. To address these issues, this paper proposes an innovative Multimodal Breast Cancer Diagnosis (MBCD) method. Specifically, we combine traditional structured feature extraction methods with the advanced EfficientNet model to extract features from ultrasound images and fuse them to improve the accuracy and comprehensiveness of breast cancer clinical diagnosis. To effectively eliminate irrelevant and redundant features and further enhance model performance, we introduce a novel hybrid feature selection method. This method utilizes SHAP feature importance analysis for filter-based feature selection and employs the Artificial Bee Colony (ABC) algorithm for wrapper-based feature selection, effectively improving feature selection efficiency and prediction accuracy. Finally, we use SHAP interpretability analysis to provide a rational explanation of the prediction results. Experimental results demonstrate that our method offers significant advantages in the clinical diagnosis of breast cancer.