Boosting medical image segmentation via conditional-synergistic convolution and lesion decoupling

Area of Science:

• Medical image analysis
• Computer vision
• Computational pathology

Background:

• Medical image segmentation is crucial for pathology assessment.
• Current deep convolutional neural networks (CNNs) struggle with complex medical images and data variations.
• Intractable cases often lead to poor performance in existing segmentation methods.

Purpose of the Study:

• To introduce a novel input-specific network, CCLDNet, for enhanced medical image segmentation.
• To address the limitations of stationary convolutions in CNNs for medical imaging.
• To improve segmentation accuracy in challenging medical image datasets.

Main Methods:

• Proposed conditional synergistic convolution (CSConv) for generating lesion-specific convolution kernels.
• Introduced a lesion decoupling strategy (LDS) to split segmentation maps into center and boundary labels.
• Utilized a transformer network backbone for dynamic modeling capabilities.

Main Results:

• CCLDNet achieved superior performance compared to state-of-the-art methods.
• Demonstrated high dice scores in polyp segmentation (89.22% on EndoScene) and skin lesion segmentation (91.15% on ISIC2018).
• The proposed CSConv can be a versatile block for various vision tasks.

Conclusions:

• CCLDNet effectively overcomes the limitations of traditional CNNs in medical image segmentation.
• The combination of CSConv and LDS significantly enhances segmentation accuracy for difficult cases.
• The developed network offers a promising advancement for automated pathology assessment and monitoring.