Unsupervised Visual Representation Learning Based on Segmentation of Geometric Pseudo-Shapes for Transformer-Based...

Area of Science:

• Medical imaging analysis
• Deep learning for computer vision
• Artificial intelligence in healthcare

Background:

• Transformer models excel in medical vision but require large labeled datasets.
• Overfitting is a challenge for transformers on smaller medical datasets.
• Unsupervised learning is crucial for overcoming data limitations in medical AI.

Purpose of the Study:

• To develop an unsupervised approach for learning semantic features in medical images.
• To train transformer models for medical image segmentation and classification without annotations.
• To introduce the Convolutional Pyramid vision Transformer (CPT) for enhanced feature learning.

Main Methods:

• Self-supervised learning using transformer-based models.
• Training models to segment numerical signals of geometric shapes on CT images.
• Developing the Convolutional Pyramid vision Transformer (CPT) with multi-kernel convolutional patch embedding and local spatial reduction.

Main Results:

• The CPT model significantly outperformed existing deep learning models on liver, pancreatic, and breast cancer datasets.
• Achieved state-of-the-art results in segmentation and classification tasks.
• Demonstrated effective unsupervised semantic feature learning on medical imaging data.

Conclusions:

• Unsupervised semantic feature learning with transformers is effective for medical imaging tasks.
• The CPT architecture offers improved multi-scale feature extraction and reduced computational cost.
• This approach reduces the need for large, annotated datasets in medical AI.