02740nas a2200277   4500000000100000000000100001008004100002260001200043653001800055653001800073653002300091653001900114653002500133100001800158700002600176700001800202700002000220700001900240700002600259245010900285856005800394300000900452490001300461520197400474022001402448       9998                            d  c03/202510aBreast Cancer10aDeep Learning10aFeature Extraction10aFeature Fusion10aFeature Optimization1 aMamuna Fatima1 aMuhammad Attique Khan1 aSaima Shaheen1 aSeifedine Kadry1 aOmar Alqahtani1 aM. Turki-Hadj Alouane00aAttention Embedded Residual Bottleneck CNN Architecture for Breast Cancer Diagnosis in Ultrasound Images  uhttps://www.ijimai.org/journal/bibcite/reference/3570  a1-110 vIn press3 aBreast cancer (BrC) stands as the predominant cancer among women, resulting in a substantial global mortality toll each year. Early detection plays a pivotal role in diminishing mortality rates. Manual diagnosis of BrC is time-intensive, intricate, and prone to errors, emphasizing the necessity for an automated system for timely detection. Various imaging methods have been investigated, underscoring the crucial need for accurate detection to prevent unwarranted treatments and biopsies. Recent years have witnessed substantial exploration and enhancement in the application of DL for efficiently processing medical images. This study aiming to create an effective and resilient DL framework for BrC detection and classification. The steps are contrast enhancement and augmentation, a hybrid CNN network ‘BrC-DeepRBNet’ is introduced that is built from scratch and
incorporates several design elements including residual blocks, bottleneck architecture, and a self-attention mechanism. This framework is employed to construct two networks, one comprising of 107 layers and the other with 149 layers. Moreover, the network capitalizes on the benefits offered by batch normalization (BN) and group normalization (GN), utilizes ReLU and leaky ReLU as activation functions, and integrates Max pooling layer into its architecture in a series of residual-bottleneck blocks. Further, for feature fusion horizontal approach is used and optimization is done using generalized normal distribution optimization (GNDO). The selected features are further classified using neural network classifiers. The introduced framework achieved the highest classification accuracy at 97.05% with publicly available BUS dataset. A detailed ablation study is presented that demonstrates the superior performance of the presented approach, surpassing various pre-trained models (i.e. AlexNet, InceptionV3, ResNet50, and ResNet101) and existing BrC detection and classification techniques.  a1989-1660