Computer Vision Analysis of Specimen Mammography to Predict Margin Status.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Intra-operative specimen mammography aids breast cancer surgery by assessing tumor margins, but struggles with microscopic positivity detection.
• Accurate margin assessment is crucial for effective breast-conserving surgery and minimizing repeat operations.

Approach:

• Developed deep learning models using specimen mammography images and pathology reports to predict margin status.
• Compared models pre-trained on radiologic images versus non-medical images for performance.
• Evaluated models using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).

Key Points:

• Models pre-trained on radiologic images generally outperformed those trained on non-medical images.
• The InceptionV3 model achieved 84% sensitivity, 42% specificity, and an AUROC of 0.71.
• Performance is comparable to human expert interpretation of specimen mammography.

Conclusions:

• Deep learning models show potential for improving intra-operative identification of positive margins in breast cancer surgery.
• Further development could lead to reduced positive margins and fewer re-operations, enhancing breast-conserving surgery outcomes.