A Multimodal Approach to Breast-Lesion Classification Using Ultrasound and Patient Metadata

The diagnosis and prognosis of breast cancer have been greatly improved by incorporating machine learning methods, especially through medical imaging analysis as well as clinical information. In this study, the potential of deep learning models for breast lesion prognosis was explored by integrating imaging features with clinical data to enhance predictive accuracy. Clinical data were analyzed using multilayer perceptron (MLP) classifiers, XGBoost, and Random Forest, while several convolutional neural network (CNN) architectures, such as ResNet optimized with Adam, DenseNet with stochastic gradient descent (SGD), and EfficientNet with RMSprop, were evaluated. The integration of imaging-based features with clinical data was found to significantly improve model performance, enabling more accurate risk stratification and the development of individualized treatment strategies. The highest validation accuracy and area under the curve (AUC) were achieved by the most effective models, highlighting the advantages of a multimodal approach. Although the study was conducted on a relatively small dataset and faced challenges such as missing data, the results suggest that these methods hold considerable promise for implementation in clinical practice.