An integrated framework for breast cancer diagnosis using adaptive DeepLabv3+ for segmentation and attention-based multiscale EfficientNetB7 for classification

Breast cancer is the primary health concern for women around the world. An advanced deep learning mechanism is proposed in this work for recognizing breast cancer disease. The proposed model initially processes image collection, where significant tomography images are gathered from online sources. Further, the original images undergo a segmentation process with the support of Adaptive Deeplabv3+ (ADV3+) for the accurate detection of subtle tumor regions. Finally, the obtained segmented images are used as input for classification, done by Optimized Multiscale EfficientNetB7 with Talking-Heads Attention (OMEB7-THA). This model is improved with Talking-Heads Attention and multiscale features, allowing it to focus on important patterns and textures across different image resolutions. For achieving the optimum performance, the primary hyperparameters in both DV3+ and OMEB7-THA are tuned appropriately using the Hunting phase Modified Frilled Lizard Optimization (HMFLO) algorithm. By combining detailed segmentation with intelligent classification, the system aims to support faster and more reliable detection of breast cancer. Finally, the effectiveness of the approach is examined by utilizing different metrics and contrasted with other baseline mechanisms. Hence, the experimental validation declares that the designed framework can provide accurate breast cancer detection results.