An improved feature extraction algorithm for robust Swin Transformer model in high-dimensional medical image...

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Medical image analysis faces challenges due to limited annotated data and image quality issues like noise and low contrast.
• High-dimensional feature extraction is crucial for complex anatomical structures in medical imaging.

Purpose of the Study:

• To propose a robust Swin Transformer architecture for high-dimensional feature extraction in medical images.
• To leverage multitask learning for improved general feature learning from limited medical data.
• To address challenges of noise, artifacts, and low contrast in medical images.

Main Methods:

• Utilized the Swin Transformer architecture, a novel deep learning model.
• Implemented an adaptive attention mechanism to dynamically adjust weights based on input quality.
• Employed an iterative transformer encoder to create a hierarchical structure for multi-scale attention.
• Trained the model using a multitask learning scheme for simultaneous analysis of various medical imaging tasks.

Main Results:

• Achieved 80.76% accuracy, 80.28% precision, 78.04% recall, and 76.46% F1-Score.
• Demonstrated effective handling of noisy and low-contrast medical images through adaptive attention.
• Showcased the model's ability to capture both local and long-range relationships in image patches at different scales.

Conclusions:

• The proposed Swin Transformer architecture offers a robust solution for high-dimensional feature extraction in medical imaging.
• The adaptive attention and hierarchical structure enable effective analysis of complex and degraded medical images.
• Multitask learning with Swin Transformer shows promise for improving generalization and performance on new medical imaging tasks.