A deep learning framework for accurate breast mass segmentation in mammographic images.
Overview
BreMSNet is a deep learning-based segmentation framework designed for automatic breast mass segmentation from mammographic images.
Accurate segmentation of breast masses remains challenging due to:
Low contrast between lesions and surrounding tissues
Large variations in lesion shape and size
Irregular and ambiguous boundaries
To address these challenges, BreMSNet introduces a multi-component encoder-decoder framework that integrates multi-scale feature learning, contextual enhancement, and boundary-aware refinement for precise lesion delineation.
Architecture
BreMSNet follows a hybrid encoder-decoder architecture consisting of four major components:
1. ResNet Encoder
A pretrained ResNet backbone is used to extract hierarchical feature representations from mammographic images.
The encoder captures:
Low-level spatial details
High-level semantic information
Multi-scale lesion representations
2. Multi-scale Hybrid Dilated (MHD) Module
The MHD module enhances contextual feature learning using hybrid dilated convolution operations.
It helps the network:
Capture lesions with different sizes
Expand receptive fields
Preserve important spatial information
3. Pyramid Refinement Module (PRM)
The Pyramid Refinement Module performs progressive feature refinement through multi-level feature aggregation.
Benefits:
Improves semantic consistency
Enhances contextual information
Strengthens lesion representation
4. Boundary-aware Feature Aggregation (BFA)
The BFA module focuses on accurate lesion boundary reconstruction.
It improves:
Boundary localization
Edge preservation
Segmentation accuracy for irregular masses
5. Decoder
The decoder progressively restores spatial resolution using:
Feature fusion
Skip connections
Progressive upsampling
The final output is a pixel-level binary segmentation mask.
