Breast cancer continues to be a major health concern worldwide. Early and accurate prediction is crucial for effective treatment and improving survival rates. A Computer Aided Diagnosis system serves as an invaluable tool for radiologists, aiming to reduce diagnostic errors and enhance the accuracy of diagnosis. These systems incorporate various processing techniques, including pre-processing, segmentation, feature extraction, and classification. Moreover, deep learning methods frequently suffer from sub optimal performance and demand substantial computational resources. This study focuses on developing an automated classification model for mammography images to aid in breast cancer diagnosis. Our proposed model initiates with noise removal using median filters, followed by the removal of the pectoral muscle in images through the Canny-edge detection method. On these processed images, we applied data augmentation using a two-point crossover technique, addressing issues of small datasets and class imbalances common in medical image analysis. The images then undergo multi-scale representation via the fourth-order complex diffusion algorithm. Feature extraction is conducted on these multi-scaled images using a Hierarchical Variational Auto-encoder. The extracted features are subsequently classified using a Support Vector Machine. Employing fourth-order complex diffusion for initial multi-scale representation significantly enhances the accuracy of feature extraction using Hierarchical Variational Auto-encoder, resulting in robust classification performance. The training process involves two different datasets like MIAS and the KAU-BCMD. Test results for the KAU-BCMD dataset include: accuracy of 99.80%, AUC of 99.30%, F1-score of 99.20%, balanced accuracy of 99.80%, and Matthews correlation coefficient of 99.20%. For the MIAS dataset, test results show accuracy of 99.30%, AUC of 99.10%, F1-score of 98.30%, balanced accuracy of 99.00%, and Matthews correlation coefficient of 99.00%. Our validation results clearly indicate that the integration of advanced techniques like fourth-order complex diffusion and Hierarchical Variational Auto-encoder in computer aided diagnosis systems not only addresses the limitations of traditional and deep learning methods but also sets a new benchmark in medical image analysis, ensuring better patient outcomes.

Deep neural networks have dramatically gained immense potential in almost every field like computer vision, natural language processing, biomedical informatics etc. Among these networks, autoencoders are popular in performing dimensionality reduction

Globally, breast cancer is the second most lethal form of cancer among women, and has high rates of incidence and mortality. Early detection is crucial in improving survival rates. As deep learning technologies continue to advance, they are increasingly being leveraged to automate pathological image classification, primarily through conventional models. Traditional models excel at processing images in which objects are clearly defined against their backgrounds, but they struggle with pathological images because the boundaries between tumor lesions and surrounding tissues are often indistinct. Hence, models that can learn the minute details in these kinds of images are very essential.We propose a hybrid approach that integrates convolutional networks with attention mechanisms to achieve high classification performance while reducing computational complexity. To enhance feature representation in convolutional neural networks, we introduced an improved convolutional block, termed the SE-Res-Conv Block, which incorporates Squeeze-and-Excitation (SE) attention mechanisms within a residual convolutional framework. The extracted features are then processed by a Conformer block, which further refines them by emphasizing the most relevant regions in the input feature map. Finally, a dense output layer is used for classification.The proposed model is evaluated on both the BACH dataset and all four magnification levels of the BreakHis dataset. Experimental results, obtained through K-fold cross-validation, indicate that the model performs optimally on 200× magnification images, achieving an average accuracy, precision, and recall of 0.97. These results underscore the model’s effectiveness and its potential applicability in medical image classification tasks.