Models-Genesis

Models-Genesis provides self-supervised pre-trained models and a unified framework for 3D medical image analysis on CT and MRI, preserving 3D anatomical information and enabling transfer learning when annotated data are limited.

Key Features:

• Ex Nihilo Creation: Models-Genesis trains models from scratch (ex nihilo), eliminating the need for manually labeled datasets.
• Self-Supervised Learning: The models employ self-supervision by using intrinsic anatomical patterns in medical images as supervision signals to learn common anatomical representations.
• Generic Framework: Models-Genesis provides generic source models that can be adapted into application-specific target models via transfer learning techniques.
• 3D Anatomical Information Preservation: The approach preserves rich 3D anatomical information in CT and MRI rather than reformulating 3D tasks into 2D.

Scientific Applications:

• Segmentation and Classification: Applied to 3D segmentation and classification tasks on CT and MRI, demonstrating improved performance relative to training from scratch or transferring from 2D ImageNet pre-trained models.
• Transfer Learning Source: Serves as a primary source of pre-trained models for transfer learning in 3D medical imaging applications.
• Limited-Annotation Environments: Suited for research and clinical scenarios with limited annotated data due to its self-supervised training strategy.

Methodology:

A unified self-supervised learning framework trains models from scratch by leveraging recurrent anatomical structures in medical images as strong supervision signals and produces source models for transfer learning.