BiLoG-Net: A Bi-Context Location-Guided Network for Breast Mass Segmentation and Malignancy Classification in Mammography

Technical Summary: BiLoG-Net for Breast Mass Segmentation and Malignancy Classification

Problem Statement
Breast cancer remains the most frequently diagnosed malignancy among women globally. Accurate detection and characterization of breast masses in mammography are hindered by several intrinsic challenges: subtle intensity variations, heterogeneous tissue densities, and indistinct lesion boundaries. Furthermore, mammographic images differ significantly from natural images; standard deep learning models designed for natural images often yield suboptimal results when applied directly to mammograms. Existing approaches often rely on multi-stage pipelines where segmentation and classification are performed sequentially, leading to error propagation. Additionally, standard Convolutional Neural Networks (CNNs) face a trade-off between maintaining high-resolution input (requiring significant computational power) and downsampling (causing information loss), which is critical for detecting small or irregularly shaped lesions.

Methodology: BiLoG-Net Architecture
The authors propose BiLoG-Net (Bi-Context Location-Guided Network), a unified, end-to-end deep learning framework designed to jointly perform breast mass segmentation and malignancy classification. The architecture follows an encoder-decoder paradigm inspired by U-Net but introduces novel components to address the specific challenges of mammographic analysis.

1. Bi-Context Location-Aware Encoder (BiLoF-Down):
   The encoder utilizes a novel block that integrates several mechanisms:

   • Fire-Based Feature Extraction: Inspired by SqueezeNet, this module uses a squeeze-and-expand strategy (1×1 convolution followed by parallel 1×1 and 3×3 convolutions) to ensure computational efficiency while capturing diverse spatial features.
   • Global and Local Feature Enhancement: The framework employs lightweight Global Feature Enhancement Modules (GFEM-Lite) and Local Feature Enhancement Modules (LFEM-Lite). GFEM-Lite uses global average pooling and channel attention to recalibrate feature responses for global context, while LFEM-Lite uses depthwise separable convolutions to capture boundary-sensitive and texture-specific details.
   • Bi-Context Fusion: Global and local features are adaptively fused using learnable scalar parameters to balance contextual awareness with spatial precision.
   • Location-Aware Gating: A spatial attention map is generated to reinforce diagnostically relevant regions (lesion-prone areas) while preserving background context, preventing the suppression of subtle low-contrast signals.

2. Shared Bottleneck and Task-Specific Heads:

   • Bottleneck: A shared bottleneck representation aggregates high-level semantic information, serving as the junction between the encoder and decoder and the primary source for malignancy prediction.
   • Decoder (BiLoF-Up): The decoder reconstructs high-resolution segmentation maps using gated skip connections. These connections selectively fuse encoder skip features with upsampled decoder features, suppressing irrelevant background noise while preserving lesion-relevant spatial regions.
   • Segmentation Head: Produces a pixel-wise probability map for mass localization.
   • Classification Head: Unlike standard global pooling, this head employs a segmentation-guided attention mechanism. It generates an attention map based on the bottleneck features and refines it using the network's own predicted segmentation mask. This ensures the classification decision is grounded in spatially meaningful, lesion-centric regions rather than global image features.

3. Training Strategy:
   The model is trained using a multi-task learning objective with a combined loss function:

   • Segmentation Loss: A weighted combination of Binary Cross-Entropy (BCE) and Dice loss to handle class imbalance and optimize spatial overlap.
   • Classification Loss: Binary Cross-Entropy for malignancy prediction.
   • Total Loss: A weighted sum of segmentation and classification losses, enabling mutual reinforcement between pixel-level localization and image-level diagnosis.

Key Contributions

• Novel Bi-Context Modeling: The introduction of a depth-wise adaptive bi-context fusion mechanism that learns the optimal balance between global and local features at each encoder stage, coupled with a residual-preserving location-aware gate to avoid suppressing subtle lesion signals.
• Unified Multi-Task Framework: A tightly coupled encoder-decoder architecture that performs segmentation and classification simultaneously under a single joint-loss framework. This design facilitates mutual reinforcement between tasks, reducing the error propagation common in multi-stage pipelines.
• Segmentation-Guided Classification: A unique mechanism where the predicted segmentation mask guides the attention mechanism in the classification branch, ensuring that malignancy assessment is focused on the specific lesion region.

Experimental Results
The framework was evaluated on two benchmark datasets: CBIS-DDSM and INBreast. All baselines (including U-Net, TransUNet, HTU-Net, and various CNN/Transformer classifiers) were re-implemented and trained under the same controlled protocol to ensure fair comparison.

• Segmentation Performance:

  • On CBIS-DDSM, BiLoG-Net achieved a Dice Similarity Coefficient (DSC) of 94.20%, IoU of 88.30%, and Accuracy of 98.95%. This outperformed the strongest baseline (HTU-Net) by 0.70% in DSC and 0.89% in IoU.
  • On INBreast, it achieved a DSC of 93.10%, IoU of 87.20%, and Accuracy of 96.05%, surpassing HTU-Net by 0.96% in DSC.
  • Qualitative results demonstrated superior boundary delineation, particularly for small and irregular lesions, compared to standard U-Net and transformer-based models.

• Classification Performance:

  • On CBIS-DDSM, the model achieved 95.20% accuracy and 97.10% AUC, outperforming the best baseline (ViT) by 1.40% in accuracy and 1.00% in AUC.
  • On INBreast, it achieved 93.60% accuracy and 96.00% AUC, outperforming ViT by up to 1.50% across metrics.

• Ablation Studies:

  • Component-wise analysis confirmed that adding Fire-based extraction, GFEM-Lite, LFEM-Lite, and the segmentation-guided classification head progressively improved performance. The full model showed cumulative DSC improvements of over 10% compared to a base U-Net.
  • Loss function ablation demonstrated that the combined BCE + Dice loss formulation yielded superior results compared to using either loss individually, balancing stable gradient flow with direct spatial overlap optimization.

Significance and Claims
The paper claims that BiLoG-Net offers a robust solution for computer-aided detection (CAD) systems by combining precise boundary delineation with reliable malignancy assessment in a single end-to-end model. The authors posit that the tight coupling of segmentation and classification allows for mutual reinforcement, producing "spatially grounded" malignancy predictions that align with clinically relevant lesion regions.

The work is presented as having strong potential for clinical settings to help radiologists prioritize suspicious cases and improve screening efficiency. However, the authors maintain a modest tone regarding clinical deployment, noting that the framework is currently validated primarily on mass lesions and requires prospective clinical and multi-reader validation before integration into routine workflows. They acknowledge limitations, including reliance on fully supervised learning, fixed input resolution (512×512), and the lack of evaluation on other abnormalities like calcifications or architectural distortions. Future work is proposed to explore semi-supervised strategies and conduct external clinical validation.