Abstract
INTRODUCTION: Early detection of breast cancer via mammography screening is essential to improve survival outcomes, particularly in low-resource settings such as the global south where diagnostic accessibility remains limited. Although Deep Neural Network (DNN) models have demonstrated high accuracy in breast cancer detection, their clinical adoption is impeded by a lack of interpretability. METHODS: To address this challenge, CorRELAX is proposed as an interpretable algorithm designed to quantify the relevance of localized regions within high-resolution mammographic images. CorRELAX evaluates the contribution of partial local information to the model's global decision-making and computes correlations between intermediate feature representations and predictions to produce global heatmaps for lesion localization. The framework utilizes a DNN trained on multi-scale crops of annotated lesions to effectively capture a spectrum of lesion sizes. RESULTS: Evaluation on the VinDr-Mammo dataset yielded F1 Scores of 0.8432 for calcifications and 0.7392 for masses. Heatmap localization accuracy was assessed using the Pointing Game metric, with CorRELAX achieving average accuracies of 0.6358 based on model predictions and 0.5602 using the correlation maps, indicating robust lesion localization capabilities. DISCUSSION: These results demonstrate that CorRELAX generates interpretable coarse-segmentation maps that enhance automated lesion detection in mammography. The improved interpretability facilitates clinically reliable decision-making and addresses a critical barrier toward the integration of AI-based methods in breast cancer screening workflows.