Breast Cancer Detection & Diagnosis in Full Field Digital Mammography with Deep Learning Approach

Abstract

Breast cancer is a significant global health concern and its early detection is critical for improving the treatment outcomes. Full-field digital mammography (FFDM) has emerged as a valuable screening tool for breast cancer, with deep learning techniques offering promising avenues for enhanced detection and diagnosis. Although, research has focused on developing deep learning models for breast cancer screening or detection on a confined scope of lesions (primarily mass & calcification), there existed a significant gap in the literature regarding the detection of multiple breast cancer lesions or abnormalities. Addressing this gap, our research introduces an innovative methodology utilizing YOLOv8 deep object detector for the detection of six different types of breast cancer lesion types: Mass, Architectural Distortion, Asymmetry, Focal Asymmetry, Suspicious Calcification and Suspicious Lymph Node. We use Vindr-Mammo dataset in our research which provides an opportunity to work upon a broad spectrum of breast cancer lesions. We also employ a novel data augmentation approach of generating artifacts with synthetic lesions to enhance the sample space. Our model demonstrated 89.3% accuracy, 0.92 F1-score and 0.72 mAP. The proposed model is a pioneer effort that effectively and consistently detects a diverse spectrum of breast cancer lesions attesting its reliability in multi-lesion breast cancer detection tasks.