Improving lesion volume measurements on digital mammograms

Area of Science:

• Medical Imaging
• Radiology
• Machine Learning in Healthcare

Background:

• Accurate breast cancer lesion volume measurement is crucial for prognosis but challenging with standard digital mammography.
• Current digital mammograms undergo vendor-specific transformations, complicating direct volume calculation from raw scanner data.

Purpose of the Study:

• To develop and validate a model for accurate breast cancer lesion volume estimation from processed digital mammograms.
• To bridge the gap between raw mammogram data and routinely used processed mammograms for volume measurement.
• To enhance the utility of mammography for breast cancer prognostication and characterization.

Main Methods:

• A physics-based algorithm for lesion volume measurement on raw mammograms was adapted for processed mammograms.
• A deep learning image-to-image translation model was employed to generate synthetic raw mammograms from processed ones.
• The method was validated on 1778 mammograms, comparing lesion volumes across different views, synthetic vs. true raw data, and mammography vs. MRI.

Main Results:

• Lesion volumes computed from different mammographic views showed high correlation (Pearson r=0.93).
• Volumes derived from synthetic raw data closely matched those from true raw data (Pearson r=0.998).
• Agreement between mammography and MRI lesion volumes was good (ICC 0.81 for consistency, 0.78 for absolute agreement).

Conclusions:

• A reliable and valid algorithm for mammographic lesion volume measurement was successfully developed.
• The algorithm demonstrates excellent reliability and good validity when compared to MRI, suggesting its clinical utility.
• This method holds promise for improving lesion characterization and breast cancer prognostication using mammography.