DCANet: Dual contextual affinity network for mass segmentation in whole mammograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate breast mass segmentation in mammograms is critical for early cancer detection.
• Current methods using mass-centered patches are inefficient and clinically unstable.
• There is a need for fully automated segmentation solutions for whole mammograms.

Purpose of the Study:

• To develop a fully automated deep learning solution for breast mass segmentation in whole mammograms.
• To address the limitations of time-consuming and unstable existing segmentation algorithms.
• To enhance computer-aided diagnosis systems with improved segmentation capabilities.

Main Methods:

• Proposed a novel Dual Contextual Affinity Network (DCANet) based on an encoder-decoder structure.
• Introduced Global-Guided Affinity Module (GAM) for long-range dependencies and homogeneous region enhancement.
• Introduced Local-Guided Affinity Module (LAM) for local semantic information and heterogeneous region differentiation.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 85.95% on the DDSM dataset.
• Achieved a Dice Similarity Coefficient (DSC) of 84.65% on the INbreast dataset.
• Demonstrated superior segmentation performance and computational efficiency compared to state-of-the-art methods.

Conclusions:

• The proposed DCANet offers a robust, fully automated approach for breast mass segmentation.
• The method provides fast and accurate diagnoses, suitable for clinical application.
• DCANet enhances the reliability and efficiency of computer-aided diagnosis in mammography.