Incremental 2D self-labelling for effective 3D medical volume segmentation with minimal annotations.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Deep learning models excel in medical image segmentation but require extensive annotated data.
• Acquiring fully annotated datasets is costly and labor-intensive, hindering advancements.
• This study addresses the challenge of training models with limited annotations.

Purpose of the Study:

• To explore the feasibility of training 2D models under severe annotation constraints.
• To optimize segmentation performance while minimizing annotation costs.
• To develop a computationally efficient method for 3D medical volume segmentation.

Main Methods:

• An incremental 2D self-labelling framework was developed for 3D medical volume segmentation.
• A 2D U-Net was trained on a single annotated slice per volume.
• The model iteratively generated and refined pseudo-labels for adjacent slices, progressively fine-tuning itself.

Main Results:

• The self-labelling approach significantly improved segmentation performance on brain MRI and liver CECT datasets.
• Dice Similarity Coefficient and Intersection over Union increased by up to 15.95% and 26.75%, respectively.
• 3D continuity was enhanced, reducing the 95th percentile Hausdorff Distance from 69.88 mm to 36.46 mm.

Conclusions:

• The proposed framework effectively leverages 2D models with self-labelling for robust 3D segmentation.
• This method achieves strong performance and coherence even with extremely sparse annotations.
• The approach offers a viable solution to reduce the annotation burden in medical imaging.