BMSMM-Net: A Bone Metastasis Segmentation Framework Based on Mamba and Multiperspective Extraction.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Deep learning for medical diagnosis

Background:

• Metastatic bone tumors severely impact patient quality of life and cancer progression.
• Current manual segmentation methods are time-consuming and subjective.
• Accurate segmentation of diverse bone lesions is crucial for improved patient outcomes.

Purpose of the Study:

• To develop a novel deep learning framework for precise and efficient segmentation of bone metastases.
• To enhance the detection of osteoblastic, osteolytic, and mixed bone lesions.
• To improve upon existing methods for bone metastasis segmentation.

Main Methods:

• Introduced BMSMM-Net, a novel segmentation framework for bone metastases.
• Integrated Bottleneck Gating Mamba (BGM) and Skip-Mamba (SKM) modules to enhance feature dependency and fusion.
• Employed a Multi-Perspective Extraction (MPE) module with varied convolutional kernels for improved sensitivity.

Main Results:

• Achieved high performance on the BM-Seg dataset with F1 scores of 91.07% for bone metastases and 95.17% for bone regions.
• Obtained mIoU scores of 83.60% for bone metastases and 90.78% for bone regions.
• Demonstrated superior segmentation accuracy and computational efficiency compared to existing models.

Conclusions:

• BMSMM-Net effectively addresses challenges in bone metastasis segmentation using BGM, SKM, and MPE modules.
• The framework offers enhanced accuracy and outperforms current advanced methods.
• BMSMM-Net's efficiency and accuracy make it suitable for clinical applications in bone metastasis detection.