A Deep Learning Model to Predict Breast Implant Texture Types Using Ultrasonography Images: Feasibility Development...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Textured breast implants are linked to breast implant-associated anaplastic large cell lymphoma (BIA-ALCL).
• Accurate identification of breast implant shell texture is crucial for BIA-ALCL diagnosis.
• Current methods like patient recall and ultrasonography have limitations in texture assessment.

Purpose of the Study:

• To evaluate the feasibility of a deep learning model for classifying breast implant shell textures.
• To assess the model's predictive performance on heterogeneous ultrasonography images.
• To establish a robust, quantitative method for implant texture analysis.

Main Methods:

• A deep learning model (ResNet-50) was trained on 19,502 retrospective breast implant ultrasound images from diverse sources (Canon, GE, public datasets).
• Model performance was validated using stratified 5-fold cross-validation and external datasets.
• Gradient-weighted Class Activation Mapping (Grad-CAM) and Shannon entropy were used for pixel contribution analysis and prediction uncertainty assessment.

Main Results:

• The deep learning model achieved high performance, with AUROC values ranging from 0.909 to 0.985 and PRAUC values from 0.748 to 0.958 across different datasets.
• The model maintained quantitative validation accuracy even when masking up to 90% of less-contributing pixels.
• Prediction uncertainty varied across image groups, being lowest for Canon (0.066) and highest for images without implants (0.777).

Conclusions:

• Deep learning models can effectively predict breast implant shell texture from ultrasonography images.
• This AI-driven approach provides a quantitative method for texture classification.
• The findings support the use of deep learning as a preliminary diagnostic tool for BIA-ALCL.