A Sisters Similarity Neural Network SSNN Model for Generalization and Detection of Mammographic Breast Cancer Lesion Abnormalities

Breast cancer remains one of the leading causes of mortality among women worldwide. Early detection through mammography significantly enhances survival rates, particularly when abnormalities are identified before metastasis. However, challenges such as tissue density, image noise, variability across mammogram devices hinder consistent diagnosis. The study proposes a robust deep learning framework to automate the detection classification of Breast abnormalities- specifically masses and calcifications. In this research a patch based preprocessing pipeline, involving articraft removal, thresholding, contrast enhancement and dynamic patch extraction resulting high quality and diverse dataset to create thousands of patches. A Novel deep Neural architecture the sisters neural network inspired by Sisters NN is designed to learn discriminative similarity features between image pairs. This approach enhances generalization performance, particular under limited data and high intraclass variability. The network achieves a validation accuracy and testing accuracy of 86.01%, with notable AUC of 0.936. The frame work has integrated an advanced model that allows to predict the unknown lesion in a unseen full scan with mAP of 0.70 and IoU of 84.5%. Additionally, segmentation is done in an enhanced way by Fuzzy c-means and Distance Transform FCDT method which has improved clustering accuracy and lesion localization even in very noisy images or ambiguous tissue regions. The Proposed model demonstrates a superior generalization performance with an accuracy of 92.3%, outperforming with existing models on mAP and AUC metrics. The Framework proposed established a foundation for scalable, best early breast cancer diagnostic tool for generalization.