BRE-CAD: AN IOT-ENABLED DEEP LEARNING FRAMEWORK FOR MAMMOGRAPHIC BREAST CANCER DETECTION AND CLINICAL DECISION SUPPORT

Abstract

Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, highlighting the urgent need for reliable computer-aided diagnosis (CAD) systems that support early screening and clinical decision-making. This paper presents BRE-CAD, an end-to-end IoT-enabled CAD framework for breast cancer detection that integrates mammographic data preprocessing, deep learning–based lesion localization, graphical clinical visualization, and database-driven diagnostic management. The proposed methodology follows a structured pipeline comprising dataset curation and annotation, model training and optimization, system integration, and longitudinal data storage. A curated subset of 5,304 mammographic images (1,768 malignant and 3,536 benign) extracted from the VinDr-Mammo dataset was processed through a dedicated data pipeline involving mass extraction, bounding box annotation, and partitioning into training (80%), validation (15%), and testing (5%) subsets. A Faster R-CNN detection model was employed to simultaneously localize suspicious regions and classify breast lesions. The trained model was embedded into a graphical user interface and connected to a PostgreSQL backend through an IoT-based architecture, enabling real-time and offline diagnostic visualization as well as secure storage of clinical records. Experimental results demonstrate stable training convergence, with validation performance achieving a recall of 0.71, precision of 0.64, and F1-score of 0.67. Image-level evaluation yielded 309 true-positive cancer detections and 735 true-negative healthy classifications. In addition to quantitative metrics, qualitative analysis conducted on a randomly selected 5% test subset confirmed reliable lesion localization, effective suppression of false detections in normal mammograms, and conservative flagging of ambiguous cases. Although the achieved performance remains moderate compared to fully supervised high-resolution benchmark systems, BRE-CAD exhibits clinically desirable recall-oriented behavior and interpretable detection outcomes. By integrating deep learning inference with IoT-based data management and clinician-in-the-loop visualization, the proposed framework provides a practical screening-oriented decision-support system aimed at assisting radiologists rather than replacing expert judgment.