Multi-dataset Collaborative Learning for Liver Tumor Segmentation.

Area of Science:

• Medical image analysis
• Deep learning applications in radiology

Background:

• Deep learning has advanced automatic biomedical image segmentation for clinical use.
• Challenges remain due to limited datasets and scarcity of labels for specific modalities.
• Magnetic resonance imaging (MRI) liver and tumor segmentation requires robust automated methods.

Purpose of the Study:

• To propose a workflow for MRI liver and tumor segmentation using external publicly available datasets.
• To enhance segmentation performance beyond a baseline model by addressing data limitations.
• To improve the robustness and efficiency of automated segmentation in clinical settings.

Main Methods:

• Utilized a workflow incorporating pseudo-labeling, unpaired image-to-image translation, and self-ensemble learning.
• Employed external, publicly available datasets to augment limited internal data.
• Established the nnU-Net model as a baseline for performance comparison.

Main Results:

• Achieved an average Dice score of 95.7% for whole liver segmentation and 72.2% for tumor segmentation.
• Obtained an average symmetric surface distance of 1.23 mm for the whole liver and 15.6 mm for the tumor.
• Demonstrated significant performance improvement over the nnU-Net baseline model.

Conclusions:

• The proposed workflow effectively enhances MRI liver and tumor segmentation accuracy.
• Leveraging external datasets is a viable strategy to overcome data scarcity in medical imaging.
• The developed methods offer more robust and efficient segmentation for clinical practice.