MC-DC: An MLP-CNN Based Dual-path Complementary Network for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning for Medical Imaging

Background:

• CNN-Transformer fusion excels in medical image segmentation but suffers from high computational cost and data demands.
• Existing methods primarily focus on encoder enhancement, neglecting decoder design in medical image segmentation.

Purpose of the Study:

• To propose a novel, computationally efficient medical image segmentation network.
• To address limitations of Transformer-based models by introducing a Multi-Layer Perceptron (MLP) based approach.
• To enhance both encoder and decoder designs for improved medical image segmentation.

Main Methods:

• Developed the MLP-CNN based dual-path complementary (MC-DC) network, replacing Transformers with MLPs.
• Introduced a dual-path complementary (DPC) module for effective multi-level feature fusion.
• Designed a dual-path decoder with cross-scale global and local feature fusion modules (CS-GF and CS-LF) and a segmentation mask feature fusion (SMFF) module.

Main Results:

• Achieved 91.69% Dice and 9.52mm ASSD for skin lesion segmentation on ISIC2018.
• Obtained 91.6% and 94.4% Dice for polyp segmentation on Kvasir-SEG and CVC-ClinicDB datasets.
• Demonstrated strong performance on lung lesion segmentation (COVID-DS36), with results up to 92.3% Dice for lung consolidation.

Conclusions:

• The proposed MC-DC network shows exceptional generalization capabilities across diverse medical imaging tasks.
• MC-DC surpasses state-of-the-art methods in segmentation accuracy while maintaining lower computational complexity.
• The MLP-CNN based approach offers a viable and efficient alternative for medical image segmentation.