MommiNet-v2: Mammographic multi-view mass identification networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current mammogram analysis often relies on single views, limiting diagnostic accuracy.
• Existing deep learning (DNN) multi-view approaches typically analyze either bilateral or ipsilateral images, not both simultaneously.
• Radiologists integrate both bilateral and ipsilateral views for optimal mammographic interpretation.

Purpose of the Study:

• To introduce MommiNet-v2, an advanced deep learning model for tri-view mammographic mass identification.
• To improve upon previous multi-view analysis by simultaneously processing bilateral and ipsilateral images, emulating radiologist practices.
• To enhance both mass detection accuracy and malignancy classification in mammograms.

Main Methods:

• Developed novel high-resolution network (HRNet)-based architectures to capture symmetry and geometry from all mammogram views.
• Implemented a multi-task learning scheme integrating Breast Imaging-Reporting and Data System (BI-RADS) and biopsy data.
• Trained a mass malignancy classification network using the multi-task learning approach.

Main Results:

• Achieved state-of-the-art results in mass detection accuracy on both the Digital Database for Screening Mammography (DDSM) and an in-house dataset.
• Demonstrated MommiNet-v2's capability to effectively aggregate information from all views for precise mass identification.
• Obtained satisfactory performance in mass malignancy classification on the in-house dataset.

Conclusions:

• MommiNet-v2 represents a significant advancement in DNN-based mammogram analysis, offering simultaneous tri-view processing.
• The model effectively emulates radiologists' reading practices, leading to superior mass detection accuracy.
• The integration of HRNet architectures and multi-task learning contributes to improved diagnostic capabilities in mammography.