Identifying Clinically Relevant Findings in Breast Cancer using Deep Learning and Feature Attribution on Local Views from High-Resolution Mammography

Diego Mellado^1,2,3,4Leondry Mayeta^1,3,4Julio Sotelo⁵Marvin Querales^4,6Eduardo Godoy^3,4,7Scarlett Lever⁸Fabian N Pardo^10,4,9Steren Chabert^3,4,8Rodrigo Salas^3,4,8*

• ¹PhD Program in Health Sciences and Engineering, UNIVERSIDAD DE VALPARAISO, Valparaiso, Chile
• ²Instituto de Tecnología para la Innovación en Salud y Bienestar (ITISB), Universidad Andrés Bello, Viña del Mar, Chile
• ³Millenium Institute for Intelligent Healthcare Engineering (iHealth), Santiago, Santiago Metropolitan Region (RM), Chile
• ⁴Center of Interdisciplinary Biomedical and Engineering Research for Health (MEDING), UNIVERSIDAD DE VALPARAISO, Valparaiso, Chile
• ⁵Departamento de Informatica, Universidad Tecnica Federico Santa Maria, Santiago, Chile
• ⁶School of Medical Technology, Faculty of Medicine, UNIVERSIDAD DE VALPARAISO, Viña del Mar, V Valparaíso Region, Chile
• ⁷Informatics Engineering School, Universidad de Valparaiso, Valparaíso, Valparaiso, Chile
• ⁸School of Biomedical Engineering, Faculty of Engineering, UNIVERSIDAD DE VALPARAISO, Valparaiso, Chile
• ⁹Escuela de Medicina, Facultad de Medicina, Universidad de Valparaiso, San Felipe, Chile
• ¹⁰Metabolic Diseases Research Laboratory (MDRL), Interdisciplinary Center for Research in Territorial Health of the Aconcagua Valley (CIISTe Aconcagua), Universidad de Valparaíso, San Felipe, Chile

Early detection of breast cancer in mammography screening is critical for improving survival rates of women, particularly in the global south, where accessibility to diagnosis is limited.Deep Neural Network models have demonstrated good performance in breast cancer detection; however, their lack of interpretability remains a significant barrier to clinical acceptance.Explainable Artificial Intelligence algorithms address this by providing insights into the model's decision-making process. However, most current methods tend to generalize over vast parts of the image, often minimizing the impact of smaller lesions, which are critical for early detection. In this work, we propose CorRELAX, an interpretable algorithm designed to assess the importance of local regions in high-resolution mammography images. By measuring how partial information from local windows contributes to the model's global decision and calculating the correlation between internal feature representations and predictions, CorRELAX reconstructs global heatmaps for lesion localization. Our approach uses a Deep Neural Network trained with multi-scale crops 1 Mellado et al.of annotated findings, accommodating a diverse range of lesion sizes. Experimental results on the VinDr-Mammo dataset show that our model achieves F1 Scores of 0.8432 for calcifications and 0.7392 for masses. Using a Pointing Game metric to evaluate the precision of the resulting heatmaps, CorRELAX reaches an average accuracy of 0.6358 using the model's predictions and 0.5602 using the resulting correlation map. These results demonstrate how interpretable coarse-segmentation maps can facilitate automated lesion detection, leading to more reliable decision-making during mammography screening.