Mammogram mass segmentation and classification based on cross-view VAE and spatial hidden factor disentanglement

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Healthcare

Background:

• Breast mass detection and characterization in mammograms are critical for early breast cancer diagnosis.
• Irregularities in mass shape, size, and texture present significant challenges for automated analysis.

Purpose of the Study:

• To develop an integrated framework for simultaneous mass segmentation and classification in mammograms.
• To leverage complementary information from different mammographic views (cranio caudal and mediolateral oblique) for improved performance.

Main Methods:

• A cross-view variational autoencoder (CV-VAE) with spatial hidden factor disentanglement was proposed.
• A feature pyramid network classifier and a U-Net-like decoder were integrated for classification and segmentation, respectively.
• The model disentangles class-specific and background factors for enhanced feature learning.

Main Results:

• The framework achieved a Dice Similarity Coefficient (DSC) of 92.46% and 93.70% for mass segmentation on public datasets.
• Classification performance reached an Area Under the ROC Curve (AUC) of 93.20% and 95.01%.
• The integrated approach demonstrated competitive results compared to state-of-the-art methods.

Conclusions:

• The proposed CV-VAE framework effectively integrates complementary information from cross-view mammograms.
• Simultaneous segmentation and classification of breast masses are achievable with high accuracy.
• This method offers a promising advancement for computer-aided diagnosis in mammography.