Artificial intelligence-based classification of breast lesion from contrast enhanced mammography: a multicenter study

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate preoperative diagnosis of breast lesions is crucial for effective treatment planning.
• Contrast-enhanced mammography (CEM) provides valuable functional information for breast lesion characterization.
• Developing advanced diagnostic tools can improve patient outcomes.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI)-based method for diagnosing breast lesions using CEM.
• To explore the biological mechanisms underlying AI-driven breast lesion classification.
• To differentiate between benign and malignant breast lesions, and further classify breast cancer subtypes.

Main Methods:

• A retrospective study involving 1430 patients who underwent CEM.
• Development of an AI model using RefineNet with a convolutional block attention module (CBAM) for feature refinement.
• Integration of AI-derived features with clinical data using an XGBoost classifier for lesion diagnosis and subtype classification.

Main Results:

• The AI model achieved an area under the curve (AUC) of 0.932 in diagnosing benign versus malignant breast lesions on an external test set.
• The model demonstrated strong performance in differentiating in situ and invasive carcinoma (AUC 0.788–0.824).
• Biological analysis linked high-risk lesions to pathways like extracellular matrix organization.

Conclusions:

• The developed AI model demonstrates high predictive performance for diagnosing breast lesions using CEM and clinical data.
• The AI approach offers a promising tool for preoperative breast lesion assessment.
• Further research into the biological underpinnings can enhance AI diagnostic capabilities.