Models Genesis: Generic Autodidactic Models for 3D Medical Image Analysis.

Area of Science:

• Deep learning
• Medical image analysis
• Computer vision

Background:

• Transfer learning from natural images to medical images is common but often requires 2D reformulation, losing vital 3D anatomical data.
• Existing 3D medical imaging tasks often compromise performance due to the loss of 3D information when adapted to 2D paradigms.

Purpose of the Study:

• To develop a novel deep learning framework for 3D medical image analysis that overcomes the limitations of 2D transfer learning.
• To introduce Generic Autodidactic Models (Models Genesis) that leverage self-supervision for learning 3D anatomical representations.

Main Methods:

• Developed a unified self-supervised learning framework using "Models Genesis" – generic, self-taught models trained ex nihilo.
• Utilized the inherent recurrent anatomy in medical images as self-supervision signals for learning common representations.
• Conducted extensive experiments on five 3D applications, including segmentation and classification tasks.

Main Results:

• Models Genesis significantly outperformed training from scratch in all tested 3D applications.
• Models Genesis consistently surpassed 2D transfer learning approaches, including fine-tuning ImageNet pre-trained models and 2D versions of Models Genesis.
• Demonstrated the critical importance of 3D anatomical information for superior performance in medical image analysis.

Conclusions:

• Self-supervised learning with Models Genesis is a highly effective paradigm for 3D medical image analysis.
• Preserving and learning from 3D anatomical information is crucial for advancing deep learning in medical imaging.
• Models Genesis offers a significant improvement over existing methods, providing a strong foundation for various 3D medical imaging tasks.