Universal conditional networks (UniCoN) for multi-age embryonic cartilage segmentation with Sparsely annotated data.

Area of Science:

• Medical imaging
• Developmental biology
• Artificial intelligence

Background:

• Osteochondrodysplasia affects 2-3% of newborns, causing bone and cartilage disorders with head malformations.
• Accurate segmentation of embryonic cartilage in 3D micro-CT images is crucial for studying these disorders in mouse models.
• Current deep learning (DL) methods struggle with accuracy and generalizability due to manual annotation burden, high costs, and complex cartilage shapes.

Purpose of the Study:

• To propose novel DL methods for accurate embryonic cartilage segmentation in 3D micro-CT images.
• To enhance DL model performance by leveraging age and spatial information for better shape representation.
• To develop robust and universal DL models for diverse medical image analysis datasets.

Main Methods:

• Developed two new DL mechanisms: one conditioned on discrete age categories and another on continuous image crop locations.
• Integrated these conditional modules into existing DL architectures (CNNs, Transformers, hybrid models).
• Evaluated performance on multi-age embryonic cartilage segmentation datasets.

Main Results:

• Achieved significant and consistent performance improvements when integrating the proposed conditional modules.
• Demonstrated an average Dice score increase of 1.7% with minimal computational overhead.
• Showcased a 7.5% improvement in performance on unseen data, indicating enhanced generalizability.

Conclusions:

• The proposed DL methods effectively leverage age and spatial information to improve embryonic cartilage segmentation accuracy.
• These novel mechanisms enhance the robustness and generalizability of DL models for medical image analysis, especially with limited annotated data.
• The approach holds potential for developing universal models capable of handling diverse datasets in developmental biology and disease research.