Mammography is the most widely used breast screening method for early detection and routine monitoring of breast cancer. Most mammogram classification or segmentation techniques are limited to simple binary classifications of the abnormalities in mammograms or classifying specific cropped regions. However, these methods encode merely the texture features of the X-ray-filmed lesions, despite the relationships and patterns. Thus, most of these models have inefficient learning capacities, hindered by supervised learning processes that require extensive annotated data, which is not feasible in the medical domain. To overcome these challenges, we model the deep texture and spatial features of every small segment of the mammogram in a complex multi-graph network. Moreover, we proposed a semi-supervised mammogram multi-graph attention network (MMGAT) to classify each element into all probabilistic categories. The attention and the dense relationships in the multi-graph network empower the proposed model with significant learning capacities, resulting in a competitive performance outperforming the recent state-of-the-art mammogram classification models. The proposed model can localise every abnormal segment of the mammogram and classify each as a malignant or benign region of a mass or calcification cancerous. The proposed MMGAT requires only 40\% of the data for training to achieve the highest AUC of 0.97 and an F1 score of 0.95 on the publically available DDSM.

Mammography is a popular diagnostic imaging procedure for detecting breast cancer at an early stage. Various deep learning (DL) approaches to breast cancer detection incur high costs and are prone to classify incorrectly. Therefore, they are not sufficiently reliable to replace existing techniques used by medical practitioners. Specifically, these DL approaches do not exploit the complex texture patterns and interactions in mammograms. These DL approaches need labelled data to enable learning, limiting the scalability of these methods because of sufficient labelled datasets. Further, these DL models lack generalisation capability to newly-synthesised patterns/textures. To address these problems, in the first instance, we design a graph model to transform the mammogram images into highly correlated multi-graphs, encoding rich structural relations and high-level texture features. Then, we consider a self-supervised learning multi-graph encoder (SSL-MG) to improve the features presentation, especially when limited labelled data is available. Finally, we design a semi-supervised mammogram multi graph convolution neural network downstream model (MMGCN) to perform multi classification of mammogram segments encoded in the multi-graph nodes. We evaluate the classification performance of MMGCN independently and with integration with SSL-MG in a model called SSL-MMGCN over several training settings. Our results reveal the efficient learning performance of SSL-MNGCN and MMGCN with 0.97 and 0.98 AUC classification accuracy in contrast to the multitask deep graph (GCN) method of Hao Du et al. (2021) with 0.81 AUC accuracy.