Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HCMA-UNet: A Hybrid CNN-Mamba UNet with Axial Self-Attention for Efficient Breast Cancer Segmentation

Haoxuan Li, Wei song, Peiwu Qin, Xi Yuan, Zhenglin Chen

TL;DR

HCMA-UNet tackles the challenge of breast cancer lesion segmentation in DCE-MRI by marrying a lightweight CNN backbone with the Multi-view Axial Self-Attention Mamba (MISM) to capture both local and global context efficiently. The MISM module combines VSSB for intra-slice feature extraction and ASA for inter-slice dependencies, with an ASC channel allocation to reduce redundancy. A novel FRLoss, consisting of Foreground Center, Positive Compactness, Boundary Aware, and Hard Negative Pixels Mining components, guides feature learning to improve discrimination and boundary accuracy. Across a private and two public DCE-MRI datasets, HCMA-UNet delivers state-of-the-art performance with significantly fewer parameters (2.87M) and lower FLOPs (126.44 GFLOPs), with FRLoss showing strong cross-architecture generalization and robustness for clinical deployment.

Abstract

Breast cancer lesion segmentation in DCE-MRI remains challenging due to heterogeneous tumor morphology and indistinct boundaries. To address these challenges, this study proposes a novel hybrid segmentation network, HCMA-UNet, for lesion segmentation of breast cancer. Our network consists of a lightweight CNN backbone and a Multi-view Axial Self-Attention Mamba (MISM) module. The MISM module integrates Visual State Space Block (VSSB) and Axial Self-Attention (ASA) mechanism, effectively reducing parameters through Asymmetric Split Channel (ASC) strategy to achieve efficient tri-directional feature extraction. Our lightweight model achieves superior performance with 2.87M parameters and 126.44 GFLOPs. A Feature-guided Region-aware loss function (FRLoss) is proposed to enhance segmentation accuracy. Extensive experiments on one private and two public DCE-MRI breast cancer datasets demonstrate that our approach achieves state-of-the-art performance while maintaining computational efficiency. FRLoss also exhibits good cross-architecture generalization capabilities. The source code is available at https://github.com/Haoxuanli-Thu/HCMA-UNet.

HCMA-UNet: A Hybrid CNN-Mamba UNet with Axial Self-Attention for Efficient Breast Cancer Segmentation

TL;DR

HCMA-UNet tackles the challenge of breast cancer lesion segmentation in DCE-MRI by marrying a lightweight CNN backbone with the Multi-view Axial Self-Attention Mamba (MISM) to capture both local and global context efficiently. The MISM module combines VSSB for intra-slice feature extraction and ASA for inter-slice dependencies, with an ASC channel allocation to reduce redundancy. A novel FRLoss, consisting of Foreground Center, Positive Compactness, Boundary Aware, and Hard Negative Pixels Mining components, guides feature learning to improve discrimination and boundary accuracy. Across a private and two public DCE-MRI datasets, HCMA-UNet delivers state-of-the-art performance with significantly fewer parameters (2.87M) and lower FLOPs (126.44 GFLOPs), with FRLoss showing strong cross-architecture generalization and robustness for clinical deployment.

Abstract

Breast cancer lesion segmentation in DCE-MRI remains challenging due to heterogeneous tumor morphology and indistinct boundaries. To address these challenges, this study proposes a novel hybrid segmentation network, HCMA-UNet, for lesion segmentation of breast cancer. Our network consists of a lightweight CNN backbone and a Multi-view Axial Self-Attention Mamba (MISM) module. The MISM module integrates Visual State Space Block (VSSB) and Axial Self-Attention (ASA) mechanism, effectively reducing parameters through Asymmetric Split Channel (ASC) strategy to achieve efficient tri-directional feature extraction. Our lightweight model achieves superior performance with 2.87M parameters and 126.44 GFLOPs. A Feature-guided Region-aware loss function (FRLoss) is proposed to enhance segmentation accuracy. Extensive experiments on one private and two public DCE-MRI breast cancer datasets demonstrate that our approach achieves state-of-the-art performance while maintaining computational efficiency. FRLoss also exhibits good cross-architecture generalization capabilities. The source code is available at https://github.com/Haoxuanli-Thu/HCMA-UNet.
Paper Structure (21 sections, 9 equations, 4 figures, 3 tables)

This paper contains 21 sections, 9 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Res Block
  4. MISM
  5. VSSB
  6. ASA
  7. Dense Block
  8. FRLoss
  9. Foreground Center
  10. Positive Compactness Loss
  11. Boundary Aware Loss
  12. Hard Negative Pixels Mining Loss
  13. Overall Loss Function
  14. Experiments and Results
  15. Private Dataset
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Overview of HCMA-UNet. (a) Loss computation module that combines Dice and Ce losses with our proposed FRLoss. (b) Output Block performs final classification based on the features from the last decoder layer. (c) Res Block for downsampling operations. (d) Dense Block for channel expansion.
  • Figure 2: Overview of MISM and its sub-modules: ASA and VSSB.
  • Figure 3: Cross-Scan Mechanism.
  • Figure 4: Visual comparison of segmentation results across ten different methods on a representative axial slice from our private dataset. Models marked with FR indicate training with FRLoss.