TriConvUNeXt: A Pure CNN-Based Lightweight Symmetrical Network for Biomedical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Biomedical image segmentation is crucial for clinical diagnosis and treatment planning.
• Self-attention networks offer high performance but are computationally expensive.
• Modified Convolutional Neural Networks (CNNs) show promise but struggle with channel interaction.

Purpose of the Study:

• To design a lightweight network block for improved feature learning in biomedical image segmentation.
• To address the limitations of modified CNNs in capturing channel interactions.
• To develop an efficient and effective CNN-based model for biomedical image segmentation.

Main Methods:

• Introduced a multi-convolutional, multi-scale convolutional network block (MSConvNeXt) integrating depthwise, deformable, and dilated CNNs.
• Incorporated channel shuffling for dynamic feature map fusion.
• Deployed the MSConvNeXt block within a U-shape symmetrical encoder-decoder network named TriConvUNeXt.

Main Results:

• TriConvUNeXt achieved a 1% higher Dice-Coefficient than UNet and TransUNet on a public benchmark dataset.
• The model demonstrated significantly lower computational costs: 81% and 97% less than baseline methods.
• Comprehensive evaluation confirmed competitive segmentation performance with reduced computational demands.

Conclusions:

• The proposed TriConvUNeXt model offers an efficient and effective solution for biomedical image segmentation.
• Lightweight CNNs, enhanced with multi-scale and channel interaction mechanisms, can rival Transformer-based approaches.
• The publicly available implementation facilitates further research and application in clinical settings.