<p>Early detection of breast cancer is one of the primary keys to survival since it is one of the significant causes of death among women. This research presents a comprehensive approach to detecting breast cancer, utilizing attention-guided multi-instance learning (MIL) enhanced with gray wolf optimization (GWO) and graph convolution aggregation (GCA). The proposed model aims to improve the classification performance and accuracy of breast cancer detection. The study includes two datasets CBIS-DDSM (a Curated Breast Imaging Subset of DDSM) and INbreast (a public database for Mammography), which contain digitized mammography images of various types, including benign, malignant, and normal. Furthermore, the approach utilizes an attention-based MIL technique to focus on the areas of interest with the highest relevance in mammogram images. Additionally, GCA helps preserve the spatial connections among the images. GWO is also utilized to optimize the model, fine-tuning the parameters for robust and efficient classification in the results. The evaluation performance on CBIS-DDSM is excellent, achieving 98.27% accuracy, 98.33% F1 score, 98.01% precision, and 98.58% recall. Likewise, the accuracy, F1 score, precision, and recall were 98.60%, 98.40%, and 98.50% on the INbreast dataset. Ablation results show that each part contributes significantly to the effective functioning of the model, with the ultimate, GWO-optimized model achieving a close approximation of perfect accuracy, 98.27%. Moreover, the findings indicate that the proposed model is more effective than conventional techniques, potentially addressing the challenge of early and accurate breast cancer detection. In the future, research efforts will be geared toward expanding the data, including other imaging modalities, and optimizing it to promote generalization and applicability to clinical cases further.</p>