Assessing the Impact of Deep Learning Backbones for Mass Detection in Breast Imaging

Breast cancer is a significant global health concern, where early detection is crucial for effective treatment and better patient outcomes. Deep learning has emerged as a promising tool for the automated analysis of breast imaging, with the potential to reduce the workload of radiologists and enhance diagnostic sensitivity. The RetinaNet architecture is especially appealing for mass detection in mammograms, merging the advantages of single-shot detectors with the feature extraction capabilities of a feature pyramid network. This paper demonstrates the critical importance of selecting the backbone feature extractor embedded in RetinaNet. It is also shown that fine-tuning pre-trained backbones on the specific domain of interest significantly contributes to enhancing performance. This study addresses a critical gap by offering an evaluation of different feature extraction backbones and training methodologies for RetinaNet models applied to high-resolution breast imaging.