Xiaoqi-Zhao-DLUT/MSNet-M2SNet

What it solves

This project provides two neural network architectures, MSNet and M2SNet, designed for high-precision medical image segmentation. It addresses the challenge of accurately identifying and isolating specific anatomical structures or pathologies (such as polyps, COVID-19 lung infections, and breast tumors) from medical scans, while maintaining low computational overhead (FLOPs).

How it works

The system uses a "Multi-scale Subtraction Network" approach. It employs both intra-layer and inter-layer multi-scale subtraction designs to capture features at different scales, which helps the model better distinguish target objects from the background. It also includes a training-free loss network to optimize the segmentation process.

Who it’s for

This project is for medical imaging researchers and developers building AI-driven diagnostic tools for tasks like polyp segmentation in colonoscopies, lung infection analysis in CT scans, and OCT layer segmentation in ophthalmology.

Highlights

• Multi-scale Subtraction: Uses specialized intra-layer and inter-layer structures to refine segmentation maps.
• Versatile Medical Application: Proven effective across diverse datasets including polyps (image and video), COVID-19 lung infections, breast ultrasound, and OCT layers.
• Computational Efficiency: Designed for low FLOPs, specifically when using the Res2Net-50 backbone.
• Award-Winning: Achieved second place in the MICCAI 2022 GOALS challenge for glaucoma OCT analysis.