Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Automated mammographic risk scoring traditionally relies on handcrafted features.
• Existing methods often have limitations in capturing complex patterns within mammograms.

Purpose of the Study:

• To develop a novel method for learning a feature hierarchy from unlabeled mammographic data.
• To apply learned features for breast density segmentation and mammographic texture scoring.

Main Methods:

• A deep learning model was developed to learn features at multiple scales from unlabeled mammograms.
• A novel sparsity regularizer was introduced to control model capacity, incorporating lifetime and population sparsity.
• The learned features were used as input for breast density segmentation and texture scoring tasks.

Main Results:

• The method achieved state-of-the-art results on three clinical datasets.
• Learned breast density scores demonstrated a strong positive correlation with manual assessments.
• Learned mammographic texture scores were found to be predictive of breast cancer risk.

Conclusions:

• The proposed method effectively learns hierarchical features for mammographic analysis.
• The approach shows strong performance in breast density segmentation and texture scoring.
• The model is versatile and applicable to various segmentation and scoring problems in medical imaging.