From CNN to Transformer: A Review of Medical Image Segmentation Models.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is vital for disease diagnosis and treatment planning.
• Deep learning, particularly U-Net and transformer-based models like TransUNet, are prevalent in this field.
• Emerging models like the Segment Anything Model (SAM) are also being explored for medical applications.

Purpose of the Study:

• To survey and analyze representative medical image segmentation models.
• To quantitatively evaluate model performance on specific medical imaging datasets.
• To discuss current challenges and future trends in the field.

Main Methods:

• Theoretical analysis of seven representative medical image segmentation models.
• Quantitative performance evaluation on Tuberculosis Chest X-rays, Ovarian Tumors, and Liver Segmentation datasets.
• Comparative study of U-Net variants, TransUNet, and SAM-based approaches.

Main Results:

• Performance metrics of surveyed models on diverse medical imaging tasks.
• Identification of strengths and weaknesses of different segmentation architectures.
• Comparative analysis highlighting the effectiveness of transformer-based and SAM approaches.

Conclusions:

• Deep learning models offer significant advancements in medical image segmentation.
• Model selection depends on specific dataset characteristics and clinical application requirements.
• Future research should focus on addressing current challenges and exploring novel segmentation techniques.