Analysis of Specimen Mammography with Artificial Intelligence to Predict Margin Status

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Breast cancer diagnostics

Background:

• Intraoperative specimen mammography aids breast cancer surgery but has low accuracy for microscopic margin positivity.
• Accurate margin assessment is crucial for effective tumor resection and patient outcomes.

Purpose of the Study:

• To develop and validate artificial intelligence (AI) models for predicting pathologic margin status using specimen mammography.
• To improve the accuracy of margin assessment in partial mastectomy procedures.

Main Methods:

• A dataset of 821 specimen mammography images with corresponding margin status was collected (2017-2020).
• AI models, including InceptionV3, were trained and validated, comparing those pretrained on radiologic versus nonmedical images.
• Performance was evaluated using sensitivity, specificity, and AUROC.

Main Results:

• AI models pretrained on radiologic images generally outperformed those pretrained on nonmedical images.
• The InceptionV3 model achieved 84% sensitivity, 42% specificity, and 0.71 AUROC.
• Model performance was notably better for invasive cancers, less dense breasts, and patients of non-white race.

Conclusions:

• Developed and internally validated AI models accurately predict margin status from specimen mammograms.
• Model performance is comparable to or better than human interpretation in existing literature.
• Further development could lead to AI-guided resection, improving cosmesis and reducing reoperations.