Network conditioning for synergistic learning on partial annotations.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Multi-organ segmentation accuracy is hindered by limited labeled data.
• Partially labeled datasets and region-based segmentation introduce inconsistencies.
• Existing methods struggle with annotation burden and background class ambiguity.

Purpose of the Study:

• To develop a framework for synergistic learning on partially labeled, region-based segmentations.
• To address inconsistencies arising from varied annotations in multi-organ segmentation tasks.
• To improve the robustness and accuracy of segmentation networks with scarce labels.

Main Methods:

• Propose CoNeMOS (Conditional Network for Multi-Organ Segmentation), a label-conditioned network.
• Utilize Feature-wise Linear Modulation (FiLM) layers for stable, efficient network conditioning.
• Employ an auxiliary network to control FiLM parameters for flexible feature extraction.

Main Results:

• Achieved state-of-the-art performance in segmenting challenging low-resolution fetal MRI data.
• Demonstrated the network's ability to learn optimal feature extraction strategies (shared vs. label-specific).
• Showcased stable training and negligible computational overhead with FiLM layers.

Conclusions:

• CoNeMOS effectively handles label inconsistencies in partially labeled region-based segmentation.
• The label-conditioning approach enables flexible synergistic learning across different organs.
• The framework offers a significant advancement for medical image segmentation with limited annotations.