Identification of Key Breast Features Using a Neural Network: Applications of Machine Learning in the Clinical Setting of Plastic Surgery

Abstract

Background: In plastic surgery, evaluation of breast symmetry is an important aspect of clinical practice. Computer programs have been developed for this purpose, but most of them require operator input. Artificial intelligence has been introduced into many aspects of medicine. In plastic surgery, automated neural networks for breast evaluation could improve quality of care. In this work, the authors evaluate the identification of breast features with an ad hoc trained neural network. Methods: An ad hoc convolutional neural network was developed on the YOLOV3 platform to detect key features of the breast that are commonly used in plastic surgery for symmetry evaluation. The program was trained with 200 frontal photographs of patients who underwent breast surgery and was tested on 47 frontal images of patients who underwent breast reconstruction after breast cancer surgery. Results: The program was able to detect key features in 97.74% of cases (boundaries of the breast in 94 of 94 cases, the nipple-areola complex in 94 of 94 cases, and the suprasternal notch in 41 of 47 cases). Mean time of detection was 0.52 seconds. Conclusions: The ad hoc neural network was successful in localizing key breast features, with a total detection rate of 97.74%. Neural networks and machine learning have the potential to improve the evaluation of breast symmetry in plastic surgery by automated and quick detection of features used by surgeons in practice. More studies and development are needed to further knowledge in this area.