Semi-supervised Label Generation for 3D Multi-modal MRI Bone Tumor Segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Oncology

Background:

• Medical image segmentation is crucial but hindered by expert annotation variability.
• Existing methods often focus on 2D or single modalities, lacking robust 3D multi-modal solutions for clinical use.
• Reliable 3D multi-modal segmentation is essential for applications like oncology.

Purpose of the Study:

• To develop a framework for generating reliable, unbiased 3D segmentation labels with minimal radiologist input.
• To reduce radiologist workload and variability in manual labeling for medical image segmentation.
• To improve the quality and reliability of labels for 3D multi-modal bone tumor segmentation.

Main Methods:

• A two-step framework combining 3D multi-modal unsupervised segmentation (feature clustering) and semi-supervised refinement.
• Generation of AI-assisted labels and comparison against traditional expert-generated labels.
• Training two 3D-Unet models for 3D multi-modal bone tumor segmentation, one with each label type.

Main Results:

• The AI-assisted labeling framework generated accurate segmentation labels with minimal expert input.
• The model trained with AI-assisted labels outperformed the baseline model in 61.67% of blind evaluations.
• Demonstrated enhancement in segmentation quality and label reliability for 3D multi-modal bone tumor segmentation.

Conclusions:

• AI-assisted labeling significantly reduces radiologist workload and improves label reliability.
• The proposed framework achieves state-of-the-art performance in 3D multi-modal segmentation.
• This approach holds potential for advancing clinical applications requiring accurate medical image segmentation.