Predictive deep learning model based on contrast-enhanced mammography for breast cancer diagnosis: a pilot study

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Contrast-enhanced mammography (CEM) interpretation relies heavily on radiologist expertise.
• Automated tools are needed to assist clinical decision-making in breast cancer diagnosis.
• Deep learning offers potential for analyzing complex medical imaging data.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting breast cancer histological diagnosis using CEM images.
• To assess the model's performance in distinguishing malignant from benign breast lesions.
• To explore the utility of AI in supporting radiologist interpretation of CEM scans.

Main Methods:

• Retrospective analysis of 106 CEM images from two centers, with histopathology as the reference standard.
• Development of an ensemble deep learning classifier using CEM images from center 1 (training/cross-validation).
• Independent external validation using CEM images from center 2, assessing metrics like accuracy, sensitivity, specificity, and AUC.

Main Results:

• The model achieved an ROC-AUC of 75% during cross-validation, with 68.7% accuracy and 68.9% sensitivity.
• External testing on center 2 data yielded an accuracy of 64.3%.
• The model demonstrated promising, though not definitive, performance in predicting malignancy from CEM images.

Conclusions:

• The developed deep learning model shows potential as a tool to support physicians in breast cancer management.
• Further research with expanded datasets is necessary to enhance model performance and clinical integration.
• AI-assisted CEM analysis could lead to more efficient and less invasive breast cancer diagnosis.