Developing a deep learning model to predict the breast implant texture types with ultrasonography image: feasibility study

Introduction Breast implants, including textured variants, have been widely used in aesthetic and reconstructive mammoplasty. However, the textured type, which is one of the shell types of breast implants, has been identified as a possible carcinogenic factor for lymphoma, specifically breast implant-associated anaplastic large cell lymphoma (BIA-ALCL). Identifying the texture type of the implant is critical to the diagnosis of BIA-ALCL. However, distinguishing the shell type can be difficult due to human memory or loss of medical history. An alternative approach is to use ultrasonography, but this method also has limitations in quantitative assessment.

Objective The objective of this study is to determine the feasibility of using a deep learning model to classify the textured shell type of breast implants and make robust predictions from ultrasonography images from heterogeneous sources.

Methods A total of 19,502 breast implant images were retrospectively collected from heterogeneous sources, including images from both Canon (D1) and GE (D2), images of ruptured implants (D3), and images without implants (D4), as well as publicly available images (D5). The Canon (D1) images were trained using Resnet-50. The performance of the model on D1 was evaluated using stratified 5-fold cross-validation. Additionally, external validation was conducted using D2 and D5. The AUROC and PRAUC were calculated based on the contribution of the pixels with Grad-CAM. To identify the significant pixels for classification, we masked the pixels that contributed less than 10%, up to a maximum of 100%. To assess model robustness to uncertainty, Shannon entropy was calculated for four image groups: Canon (D1), GE (D2), ruptured implant (D3), and without implants (D5).

Result The deep learning model achieved an average AUROC of 0.98 and a PRAUC of 0.88 in the Canon dataset (D1). For images captured with GE (D2), the model achieved an AUROC of 0.985 and a PRAUC of 0.748. Additionally, the model predicted an AUROC of 0.909 and a PRAUC of 0.958 for a dataset available online. For quantitative validation, this model maintained PRAUC up to 90% masking of less contributing pixels, and the remnant pixels located in breast shell layers. Furthermore, the prediction uncertainty increased in the following order: Canon (D1), GE (D2), ruptured implant (D3), no implant (D5) (0.066; 0072; 0.371; 0.777, respectively).

Conclusion We have demonstrated the feasibility of using deep learning to predict the shell types of breast implants. With this approach, the textured shell types of breast implants can be quantified, supporting the first step in the diagnosis of BIA-ALCL.