Multi-scale feature fusion for breast cancer detection using circular dilated convolutional transformer optimized by enhanced wombat algorithm

Manual diagnosis of breast cancer by pathologists using biopsy tissue images is expensive, time-consuming, and subject to inter-expert variability. Computer-Aided Diagnosis (CAD) systems have made possible the faster, more reproducible, and more accurate detection of breast cancer. This paper proposes an innovative diagnostic system based on a Circular Dilated Convolutional Transformer (CDCT) for the diagnosis of breast cancer using two publicly available histopathological image datasets containing 7909 and 400 microscopic images. It follows the workflow: images are pre-processed using the Adaptive Morphological Wavelet Perona–Malik Filter Algorithm (AMWPMFA), which efficiently removes noise while maintaining the most important structural features. The Rotation-Invariant Attention Network (RIAN) is employed in the feature extraction step to give attention to the salient regions in images irrespective of their orientations. Features from the attention network are fused and provided to CDCT, which extracts both local and global contextual features because of its circular and dilated convolution structure. The Enhanced Wombat Optimization Algorithm (EWOA) is implemented for the optimization of model parameters. The proposed CDCT framework has achieved 99.92% binary classification accuracy and 99.91% multi-class classification accuracy, clearly indicating its accuracy and dependability as a very effective and powerful diagnostic technique for clinical diagnosis of breast cancer.