RMIS-Net: a fast medical image segmentation network based on multilayer perceptron.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation is crucial for diagnostics and treatment planning.
• Existing methods face challenges with computational complexity and efficiency.
• Need for accurate and computationally efficient segmentation networks.

Purpose of the Study:

• To introduce RMIS-Net, a novel lightweight network for efficient medical image segmentation.
• To enhance segmentation accuracy while reducing computational load.
• To address limitations of existing segmentation methods.

Main Methods:

• Developed RMIS-Net with a convolutional layer, shift-based fully connected layer, and tokenized multilayer perceptron.
• Incorporated layer normalization, dropout, GELU activation, residual connections, Dice loss, and bilinear interpolation.
• Evaluated on cellular structure (2018 Data Science Bowl) and lesion segmentation (ISIC-2018) datasets.

Main Results:

• Achieved state-of-the-art performance with an average F1-score of 0.91 and mean IoU of 0.82.
• Demonstrated significant efficiency: 27x parameter compression, 10x faster inference, and 53x less computational complexity.
• Achieved 0.03s per image inference time.

Conclusions:

• RMIS-Net provides a highly efficient and accurate solution for medical image segmentation.
• The network sets new benchmarks for performance and computational efficiency in medical image analysis.
• Potential for widespread clinical application in diagnostics and surgical navigation.