ADVANCED COMPUTATIONAL IDENTIFICATION OF BREAST CANCER BIOMARKERS: SYNERGIZING METAHEURISTIC OPTIMIZATION WITH MULTI-MODAL VISION TRANSFORMERS

Abstract

Medical image processing to identify breast cancer is particularly challenging since the imprecision can be a result of a number of varying factors such as the specific modality, the deep learning model, or low interpretability of the model, which can be problematic for real-world consequence setups. Mammography and ultrasound imaging are the primary methods and most frequently utilized imaging tools for breast cancer screening. The most difficult problem in using both technologies is drawing the right conclusions. In this paper, we introduce MMViT-Net, a multi-modality vision transformer ensemble for breast cancer imaging and a metaheuristic vision transformer. MMViT-Net uses both mammography and ultrasound images for breast cancer screening. Ensembling the distance variations a dual modality adaptive preprocessing and attention U-Net lesion segmentation will be able to localize the tumor regions and suppress the irrelevant background regions. Each hybrid modality of the encoders, egg, and vision of the transformers are coupled to capture finer and more granular attributes while elongated global contextual encoders portray the vastness of the context and the encoders. Robust cross-attention modality-aware feature fusion is proposed to improve the internal cooperation of the various modalities. The model hyperparameters are automatically adjusted by a fusion of the Grey Wolf Optimizer and Harris Hawks Optimization. This McDo analysis algorithm is utilized to optimize the trade-off between reducing false positives and increasing the positive diagnostical certainties. Results from the analyses performed on the ultrasound database, BUSI, and the mammography database, CBIS DDSM, show imposing evidence of strong performance, cross-dataset generalization, and demonstrable improvements in construction to many existing CNN and CNN-Transformer methods. MMViT-Net is positioned as a solid and clinically viable model for AI-assisted screening and diagnosis of breast cancer. This is due to the assurance of model interpretability and the support of transparent decision-making by the highlighting of clinically relevant lesion areas through Grad-CAM visual explanations.