CeLNet: a correlation-enhanced lightweight network for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Convolutional neural networks (CNNs) excel at medical image segmentation but often require complex architectures.
• High complexity leads to increased parameters and training challenges, while lightweight models may sacrifice contextual information.
• Balancing segmentation accuracy and computational efficiency is crucial for practical applications.

Purpose of the Study:

• To introduce a novel correlation-enhanced lightweight network (CeLNet) for medical image segmentation.
• To improve the balance between model efficiency (parameters, computation) and segmentation performance.
• To enhance feature extraction and contextual information utilization in lightweight models.

Main Methods:

• Proposed a siamese structure for weight sharing and parameter reduction.
• Introduced a point-depth convolution parallel block (PDP Block) for efficient feature extraction.
• Designed a relation module with global and local attention to capture inter-slice feature correlations.

Main Results:

• Achieved excellent segmentation performance on LiTS2017, MM-WHS, and ISIC2018 datasets.
• The model utilizes only 5.18M parameters.
• Reported Dice Similarity Coefficients (DSC) of 0.9233 (LiTS2017), 0.7895 (MM-WHS), and 0.8401 (ISIC2018).

Conclusions:

• CeLNet demonstrates state-of-the-art segmentation performance.
• The proposed lightweight network effectively enhances feature extraction and contextual understanding.
• CeLNet offers a practical solution for medical image segmentation with limited resources.