MetaBreastAI

Breast cancer metastasis is one of the most consequential clinical problems in breast cancer, as it is responsible for most of the breast cancer attributable deaths in women globally. Timely detection of metastasis progression is crucial for improving therapies and enhancing patient survival. Although deep learning approaches, especially convolutional neural networks (CNNs), still fail to model global dependencies, they cannot work under weak supervision, and unable to generate interpretable predictions. Transformer-based models, on the other hand, provide a deeper contextual knowledge; but they are insensitive to local patterns and require large datasets. To overcome these issues, we present MetaBreastAI, an explainable dual-stream deep learning framework comprising a CNN branch fused with the convolutional block attention module and a Transformer stream. Each of these parallel branches captures spatial and contextual features, respectively, which are fused by a multi-instance learning approach for partially supervised classification. The proposed model has been tested on benchmark computed tomography scan dataset, it enables identifying one of the distinct features of MetaBreastAI and apply feature attribution methods to both visual and quantitative experiments, demonstrating that MetaBreastAI achieves better performance (macro-averaged F1-score: 91.7%; area under the curve: 96.1%) compared to each branch independently, and hybrid baseline models. By highlighting the lesion’s location, the heatmap supports interpretability. Due to the explainable detection of lesions in metastatic cases, our hybrid model provides a scalable and clinically feasible strategy through overcoming the downsides of earlier research, then it improves the reliability of AI-assisted techniques in medical decision-making.