A novel dual-network architecture for mixed-supervised medical image segmentation.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning models for medical image segmentation demand extensive, precise annotations, which are costly and time-consuming to create.
• Mixed-supervised learning, utilizing both dense and weak annotations (e.g., bounding boxes), offers a potential solution to reduce annotation burden.

Purpose of the Study:

• To propose a novel network architecture, the Mixed-Supervised Dual-Network (MSDN), for medical image segmentation using mixed-supervised data.
• To enhance information transfer between segmentation and detection tasks within the MSDN framework.

Main Methods:

• Developed a dual-network architecture (MSDN) with separate segmentation and detection networks.
• Incorporated connection modules with a 'Squeeze and Excitation' variant for effective feature sharing between networks.
• Designed a flexible and trainable feature sharing mechanism, distinct from shared backbone models.

Main Results:

• The MSDN model demonstrated superior performance compared to multiple baseline methods across four medical image segmentation datasets.
• The proposed connection modules effectively facilitated information transfer, boosting segmentation task performance.
• The flexible feature sharing approach contributed to the model's effectiveness and stability.

Conclusions:

• The Mixed-Supervised Dual-Network (MSDN) presents a highly effective and stable approach for medical image segmentation using mixed-supervised learning.
• MSDN offers a significant advancement over existing methods by enabling efficient utilization of weakly annotated data.
• This architecture provides a promising direction for reducing the reliance on fully annotated datasets in medical imaging AI.