Comparative analysis of deep learning methods for lesion detection on full screening mammography

Breast cancer is the most prevalent type of cancer in women. Although mammography is used as the main imaging modality for the diagnosis, robust lesion detection in mammography images is a challenging task, due to the poor contrast of the lesion boundaries and the widely diverse sizes and shapes of the lesions. Deep Learning techniques have been explored to facilitate automatic diagnosis and have produced outstanding outcomes when used for different medical challenges. This study provides a benchmark for breast lesion detection in mammography images. Five state-of-art methods were evaluated on 1592 mammograms from a publicly available dataset (CBIS-DDSM) and compared considering the following seven metrics: i) mean Average Precision (mAP); ii) intersection over union; iii) precision; iv) recall; v) True Positive Rate (TPR); and vi) false positive per image. The CenterNet, YOLOv5, Faster-R-CNN, EfficientDet, and RetinaNet architectures were trained with a combination of the L1 localization loss and L2 localization loss. Despite all evaluated networks having mAP ratings greater than 60%, two managed to stand out among the evaluated networks. In general, the results demonstrate the efficiency of the model CenterNet with Hourglass-104 as its backbone and the model YOLOv5, achieving mAP scores of 70.71% and 69.36%, and TPR scores of 96.10% and 92.19%, respectively, outperforming the state-of-the-art models.