Improved Loss Function for Mass Segmentation in Mammography Images Using Density and Mass Size.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology Diagnostics

Background:

• Breast cancer mass segmentation is crucial for diagnosis, providing location, size, and border details.
• Challenges in segmentation include pixel class imbalance and variations in mass appearance and size.
• Existing loss function formulations for imbalance show improvements but lack comprehensive solutions.

Purpose of the Study:

• To propose a novel perspective on loss calculation for improved breast cancer mass segmentation.
• To develop a hybrid loss function integrating sample-level and region-level information.
• To enhance segmentation performance by considering mass size and density.

Main Methods:

• Introduced a hybrid loss setting combining binary segmentation loss with sample-level and region-level information.
• Developed two loss function variations incorporating mass size and density.
• Proposed a variant enhancing focal loss using mass size and density.

Main Results:

• The proposed hybrid loss function significantly improved breast cancer mass segmentation performance.
• The method demonstrated superior results compared to baseline and state-of-the-art approaches.
• Effectiveness validated on benchmark datasets: CBIS-DDSM and INbreast.

Conclusions:

• The novel hybrid loss function offers a more comprehensive approach to addressing segmentation challenges.
• Incorporating sample and region-level information, along with mass characteristics, enhances segmentation accuracy.
• This method shows promise for improving automated breast cancer detection and diagnosis.