Presegmenter Cascaded Framework for Mammogram Mass Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate breast mass segmentation in mammograms is crucial for early cancer diagnosis.
• Existing deep learning models struggle with false positives, false negatives, and end-to-end segmentation challenges.

Purpose of the Study:

• To develop a novel two-stage, end-to-end cascaded deep learning framework for enhanced breast mass segmentation.
• To improve segmentation accuracy and reduce false negatives in mammogram analysis.

Main Methods:

• A two-stage cascaded framework utilizing a saliency map to guide segmentation.
• Integration of presegmenter attention (PSA) blocks for dynamic focus on informative regions.
• Comparative analysis using Attention U-net, DeepLabV3+, and Swin transformer U-net on INbreast, CSAW-S, and DMID datasets.

Main Results:

• The proposed cascaded framework significantly improved segmentation performance across all datasets.
• Demonstrated improvements in Dice scores (up to 6%) and reductions in false negatives (up to 19%).
• Validated effectiveness with multiple state-of-the-art segmentation models.

Conclusions:

• The novel cascaded framework enhances breast mass segmentation accuracy and reduces critical false negatives.
• This approach offers a robust solution for improving mammogram analysis and supporting early cancer diagnosis.
• The framework's adaptability makes it a valuable tool for various medical image segmentation tasks.