PIS-Net: Efficient Medical Image Segmentation Network with Multivariate Downsampling for Point-of-Care

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Point-of-care (PoC) imaging is gaining popularity over traditional methods.
• Instant image segmentation is crucial for PoC diagnosis but faces challenges like image distortion and low resolution.
• Efficient feature representation is needed for effective instant image segmentation.

Purpose of the Study:

• To propose a new feature representation for instant image segmentation.
• To address image distortion and low resolution in PoC medical images.
• To enhance the generalization and performance of segmentation models.

Main Methods:

• Developed a Diagonal-Axial Multi-Layer Perceptron for global feature correlation.
• Introduced a multi-scale feature fusion to integrate linear and nonlinear features.
• Constructed dynamic residual spatial pyramid pooling to mitigate image distortion.

Main Results:

• The proposed method achieves superior performance in instant image segmentation.
• Demonstrated an average improvement of 1.31% in Dice score on BUSI, ISIC2018, and MoNuSeg datasets.
• The new feature representation offers minimal parameters and lower computational complexity.

Conclusions:

• The proposed strategy effectively improves instant image segmentation for PoC diagnostics.
• The novel feature representation and dynamic pooling enhance model generalization and accuracy.
• This work contributes to more reliable and efficient medical image analysis in portable diagnostic devices.