3D Image Segmentation With Sparse Annotation by Self-Training and Internal Registration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate anatomical image segmentation is crucial for medical planning and analysis.
• Convolutional Neural Networks (CNNs) excel at 3D image segmentation but require extensive annotated data.
• Acquiring fully annotated 3D medical datasets is challenging due to the time-consuming nature of manual segmentation.

Purpose of the Study:

• To develop an effective method for training 3D CNNs for medical image segmentation using sparse annotations.
• To reduce the burden of manual annotation by utilizing only one 2D slice per 3D image.
• To propose a self-training framework that leverages both 2D slice propagation and 3D volumetric information.

Main Methods:

• A self-training framework alternating between pseudo-label assignment and network updates.
• A 2D registration-based method for propagating labels between adjacent slices.
• A 3D U-Net architecture to leverage volumetric information for enhanced segmentation.

Main Results:

• The proposed method effectively trains 3D segmentation networks using only one annotated slice per 3D image.
• Cooperation between 2D registration and 3D segmentation yields accurate pseudo-labels.
• Achieved higher performance compared to other weakly supervised methods and approached fully supervised results on abdominal organ segmentation datasets (CHAOS and Visceral).

Conclusions:

• The self-training framework significantly alleviates the need for extensive manual annotation in 3D medical image segmentation.
• This approach demonstrates the feasibility of achieving high-performance segmentation with minimal supervision.
• The method offers a practical solution for training deep learning models on limited annotated volumetric medical data.