Progressive Dual Priori Network for Generalized Breast Tumor Segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Oncology Imaging

Background:

• Accurate breast tumor segmentation in dynamic enhanced magnetic resonance imaging (DCE-MRI) is crucial for diagnosis and treatment planning.
• Existing models often struggle with generalization across different centers and segmentation of challenging tumors (small size, low contrast, irregular shapes).
• Improving model generalization and performance on diverse tumor characteristics remains a significant challenge in automated medical image analysis.

Purpose of the Study:

• To develop a novel network, PDPNet, for robust breast tumor segmentation from multi-center DCE-MRI data.
• To enhance segmentation performance for small, low-contrast, and irregularly shaped breast tumors.
• To improve the generalization ability of breast tumor segmentation models across different acquisition centers.

Main Methods:

• Proposed a progressive dual priori network (PDPNet) incorporating a coarse-segmentation based localization module.
• Implemented progressive refinement using weak semantic prior and cross-scale correlation prior knowledge.
• Validated PDPNet against state-of-the-art methods on multi-center DCE-MRI datasets.

Main Results:

• PDPNet achieved significant improvements in Dice Similarity Coefficient (DSC) by at least 5.13% and Hausdorff Distance 95th percentile (HD95) by at least 7.58% compared to the suboptimal method on multi-center test sets.
• Ablation studies confirmed the localization module's effectiveness in reducing normal tissue influence and enhancing generalization.
• Weak semantic priors aided in focusing on tumor regions, preventing missed detections of small or low-contrast tumors, while cross-scale correlation priors improved segmentation of irregular shapes.

Conclusions:

• The proposed PDPNet effectively segments breast tumors from multi-center DCE-MRI, outperforming existing methods.
• The integrated localization and progressive refinement strategy, leveraging dual prior knowledge, successfully addresses challenges posed by tumor size, contrast, shape, and multi-center variations.
• PDPNet demonstrates enhanced generalization ability and improved segmentation accuracy, offering a promising tool for clinical applications.