Deep Learning-Based Model for Breast Implant Classification in Ultrasonography: A Multi-Institutional Model Development and Validation Study.

Abstract

The increasing prevalence of implant-based breast surgeries highlights a critical gap in patient knowledge regarding implant information, exacerbated by inadequate record-keeping and emerging safety concerns. The authors of this study address the need for reliable implant identification methods by developing a deep learning model capable of classifying breast implants using ultrasound images. Retrospective data of 28,712 breast ultrasound PNG files from 4136 breast implants in 2580 patients obtained from multiple institutions were utilized to train and validate this model. The findings demonstrate that the deep learning model achieved high diagnostic accuracy, with a balanced accuracy of 0.893 for manufacturer classification and 0.971 for implant texture classification in external test datasets. The model's performance was enhanced by employing Gradient-Weighted Class Activation Mapping (Grad-CAM) for interpretability. By automating the identification process, this tool alleviates the reliance on specialized training among plastic surgeons regarding breast ultrasound, streamlining patient care. Despite limitations, the model shows promise for improving clinical workflows and patient outcomes.