Modèles graphiques neuronaux des images mammographiques pour une détection explicable du cancer du sein

Breast cancer is a significant global health burden, with an estimated 1.9 million new cases diagnosed in 2022, representing 12.7% of all cancer cases. It is the leading cause of cancerrelated deaths among women, showing the importance of a screening strategy, allowing it to significantly lower the mortality rate. However, mammography’ findings rely on precise analysis ofshape and texture patterns in the image to discern healthy tissues from suspicious ones. Moreover, as mammography is acquired with complementary and high-resolution views, analyzing these images can be time-consuming, leading to fatigue and burnoutamong radiologists. To mitigate these issues, Computer-Aided Diagnostic (CAD) tools have been developed to assist radiologists. Deep learning techniques showed the potential to overcome the limitations of traditional image processing algorithms. We chose a deep learning approach using graph representation of images to provide explainable and relevant results, despite challenges such as precise image annotations, data heterogeneity, and highresolution images. Our method involves segmenting the image into clusters based on textural information, creating a graph representation of the image, and using a Graph Neural Network (GNN) to analyze and localize/classify suspicious regions. We also introduce a graph representation of dual-view mammography complementarity to consider similar information from multiple images. Our methods are interconnected and form part of a final solution to evaluate breast malignancy, contributing to the current state of mammography detection and classification tasks and opening the way for future research in the field.