Lightweight semantic segmentation network for tumor cell nuclei and skin lesion

Area of Science:

• Medical Image Analysis
• Computational Biology
• Artificial Intelligence

Background:

• Accurate semantic segmentation of tumor cell nuclei and skin lesions is crucial in medical imaging.
• Variations in lesion morphology and cell types challenge existing segmentation models.
• Increasing network depth leads to high computational costs and parameter sizes, hindering practical use.

Purpose of the Study:

• To develop a fast and accurate segmentation network for medical images.
• To address the challenges of varying lesion appearances and cell types.
• To reduce the computational complexity and parameter size of deep learning models.

Main Methods:

• Proposed MD-UNet architecture integrating Tokenized Multi-Layer Perceptron (MLP) modules, attention mechanisms, and Inception structures.
• Utilized tokenized MLP modules for feature labeling and projection, reducing computational load.
• Incorporated Depthwise Attention blocks and Multi-layer Feature Extraction modules to refine feature representation and fusion.

Main Results:

• MD-UNet demonstrated superior performance in segmentation tasks on both the ISIC2018 and PanNuke datasets.
• The proposed architecture effectively handles variations in medical image data.
• Achieved high accuracy and robustness in segmenting tumor cell nuclei and skin lesions.

Conclusions:

• MD-UNet provides an effective solution for fast and accurate medical image segmentation.
• The integration of novel modules enhances network performance and efficiency.
• The model shows significant potential for clinical applications in oncology and dermatology.