LungNet: Leveraging state-space models with SE-enhanced skip connections for precise CT-based lung lesion...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer remains a leading cause of cancer death globally, necessitating advanced diagnostic tools.
• Computed tomography (CT) imaging is crucial for lung cancer detection, but accurate segmentation of complex lesions is challenging.
• Existing deep learning models like CNNs and Transformers have limitations in capturing long-range dependencies or are computationally expensive.

Purpose of the Study:

• To develop a novel deep learning framework for accurate lung lesion segmentation in CT imaging.
• To overcome the limitations of current methods in capturing long-range dependencies and handling feature redundancy.
• To improve the precision of lung lesion boundary delineation for enhanced cancer diagnosis and management.

Main Methods:

• Integration of Mamba state-space models with an improved UNet architecture.
• Embedding Squeeze-and-Excitation networks into skip connections to mitigate feature redundancy.
• Introduction of auxiliary losses to preserve shallow features and capture fine-grained details of lesions with varied morphologies.

Main Results:

• The proposed framework effectively models global context and maintains linear computational efficiency.
• Achieved state-of-the-art segmentation accuracy on multiple datasets, outperforming existing methods.
• Demonstrated precise delineation of lung lesion boundaries, accommodating variations in size, shape, and spatial distribution.

Conclusions:

• The novel deep learning framework offers a robust solution for complex lung lesion segmentation.
• This advancement holds significant promise for improving early lung cancer detection, disease monitoring, and treatment planning.
• The findings have implications for developing more effective and potentially cost-efficient diagnostic tools in oncology and the insurance sector.