Efficient AI-Based breast cancer diagnosis using a pruned Adaptive dilated residual attention network with transformer network for Low-Computational environments

The primary goal of this framework is to enhance the survival rate of breast cancer patients through early and accurate identification using an artificial intelligence-based diagnosis approach. This approach reduces computational costs, memory usage, and data requirements, making the system ideal for low-power computing. The methodology involves a multi-stage deep learning framework starting with the collection of mammogram images from three diverse datasets. Region of Interest (ROI) detection is performed using a novel Trans-YoloUnet++ (TYUNet++) model. Classification is then conducted using the Adaptive Dilated Residual Attention Network-Spatial Pyramid Pooling (ADRAN-SPP) integrated with pruning and compressive filters to reduce redundant parameters. Furthermore, hyperparameters such as activation functions and learning rates are optimized using the Refined Fitness-based Addax Optimization Algorithm (RF-AOA). Finally, the developed model offers the classified outcome. The effectiveness of the designed model is confirmed in the validation. The proposed model demonstrates strong performance, achieving an ROI extraction accuracy of 91.6%, with Intersection over Union (IoU) and Dice scores of 92% and 93%, respectively. In addition, the classification accuracy reaches 97.8%, confirming the effectiveness of the proposed approach. This framework concludes that the integration of pruning and compressive filters with optimized deep learning architectures provides a robust, high-accuracy solution for breast cancer diagnosis.