SAP-cGAN: Adversarial learning for breast mass segmentation in digital mammogram based on superpixel average pooling

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast mass segmentation is crucial for computer-aided diagnosis and treatment planning in breast cancer.
• Challenges in segmentation include low contrast, irregular shapes, and fuzzy boundaries of masses.

Purpose of the Study:

• To propose an improved mammography mass segmentation model for enhanced diagnostic accuracy.
• To address the limitations of existing segmentation methods in challenging cases.

Main Methods:

• Developed SAP-cGAN, an enhanced conditional generative adversarial network (cGAN).
• Incorporated a superpixel average pooling layer for improved boundary segmentation.
• Utilized a multiscale input strategy for learning scale-invariant features and robustness.

Main Results:

• Achieved high performance on CBIS-DDSM (Dice: 93.37%, Jaccard: 87.57%) and INbreast (Dice: 91.54%, Jaccard: 84.40%) datasets.
• Demonstrated superior performance compared to current state-of-the-art methods.
• Superpixel average pooling and multiscale input improved Dice and Jaccard scores by 7.8% and 12.79% over the original cGAN.

Conclusions:

• Adversarial learning, combined with superpixel pooling and multiscale inputs, enhances breast mass segmentation realism and performance.
• The proposed SAP-cGAN model offers a significant advancement in automated breast mass segmentation.