Automated Registration for Dual-View X-Ray Mammography Using Convolutional Neural Networks

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Image registration algorithms

Background:

• Accurate lesion localization in mammography is crucial for diagnosis.
• Standard mammography involves multiple views (CC, MLO) requiring image registration.
• Existing registration methods can be time-consuming or lack accuracy.

Purpose of the Study:

• To develop automated registration algorithms for 2D X-ray mammographic images.
• To accurately map lesions between craniocaudal (CC) and mediolateral oblique (MLO) views.
• To enhance lesion co-localization for improved diagnostic accuracy.

Main Methods:

• A fully convolutional neural network (CNN) was employed to generate pixel-level deformation fields.
• The CNN creates a mapping between lesions in CC and MLO views.
• Novel distance-based regularization was implemented to enhance performance.

Main Results:

• The developed algorithms were tested on real and synthetic mammographic images.
• Performance was evaluated across various factors like image resolution, breast density, and lesion characteristics.
• The CNN-based approach outperformed existing state-of-the-art registration techniques.

Conclusions:

• The proposed automated methods provide a tool for co-locating lesions between CC and MLO views.
• The algorithms demonstrate robust performance, even in challenging cases.
• This technology can assist clinicians in establishing lesion correspondence quickly and accurately, improving diagnostic capability.