Imaging phenotype evaluation from digital breast tomosynthesis data: A preliminary study.

Area of Science:

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

Background:

• Digital breast tomosynthesis (DBT) is a key imaging tool for breast cancer diagnosis.
• Current methods require further refinement for detailed tissue pattern characterization.
• Deep learning offers potential for advanced image analysis in mammography.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for characterizing breast tissue patterns in DBT data.
• To classify tissue samples into malignant, benign, and normal categories.
• To perform a more intricate classification of specific tissue abnormalities like masses and architectural distortions.

Main Methods:

• A dataset of 5388 2D image patches from DBT studies was curated.
• A patch classifier was trained for two classification scenarios: malignant-benign-normal and detailed abnormality classification.
• Transfer learning was utilized, initializing model weights from a pre-trained Globally-Aware Multiple Instance Classifier.

Main Results:

• High recall and specificity were achieved for malignant, benign, and normal tissue classification.
• Detailed classification of masses and architectural distortions also showed promising recall and specificity values.
• Some confusion was observed between benign architectural distortion and benign mass classifications.

Conclusions:

• The developed deep learning phenotype classifier enhances the assessment of DBT images.
• Combining this classifier with standard malignant-benign-normal classification provides more detailed diagnostic information.
• This approach has the potential to improve the accuracy and granularity of breast cancer diagnosis using DBT.