Realizing Nipple in Profile Recognition and Nipple Detection Using a Single Classification

Nipple in Profile (NIP) is one of the positioned criteria for mammography quality evaluation. The current evaluation work relies on manual reviewing, which is subjective and time-consuming. This paper introduces a novel NIP recognition based on a convolutional neural network (CNN) without referring to prior information, which outputs an objective recognition result in real-time. Amazingly, the added value associated with the proposed method is nipple detection which analyzes the extracted gradient-weighted class activation mapping (Grad-CAM) based on three-category classification. In this work, the defined categories are Nipple in Profile, Nipple not in Profile, and non-nipple. For the NIP recognition, we only exploit the comparison between the Grad-CAMs of Nipple in Profile and Nipple not in Profile. Our method is more creative than the typical classification approaches to realize multiple tasks with a single classification. The experiment results demonstrate that the three-category classification with Densenet-121 achieves better performance in both NIP recognition and nipple detection.